JP4056546B2 - Search device, search method and database for compound nouns input by speech - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently retrieve a retrieval word input in voice by a user in a short period of time when compound nouns of the number exceeding the retrieval word number processable in interaction processing real time are registered in a retrieval database, and compound nouns having a large number of mutually closely resembled candidates are registered in the retrieval database. <P>SOLUTION: When it is decided as a result of recognition of a starting beat that the beat is a starting beat that the user has input with high possibility, beats which connects with the possible starting beat are preferentially recognized in the decreasing order of use frequencies to extract a 2nd possible beat, and the recognition processing is repeated until the final beat constituting retrieval information; when the recognition processing of all beats is completed, total scores are calculated for each of output pieces of retrieval information and when a total score meets a condition under which the user can specify the retrieval information only by confirming it, interaction with the user is carried out. <P>COPYRIGHT: (C)2007,JPO&amp;INPIT

Description

  The present invention relates to a compound noun search device and a search method in which a compound noun composed of a plurality of single nouns is inputted by a user and the compound noun inputted by voice is specified.

That is, in the present invention, the search term to be searched is a compound noun composed of a plurality of single nouns, and when a large number of search terms are held in the search database, the speech recognition process and the search are performed. The present invention relates to an apparatus and method for specifying a search term without causing the user to feel the time required for the above and without causing an unnatural conversation due to a defect in voice recognition processing or the like.

  In a conventional database search process using a speech recognition device, a speech inputted by a user is compared with a recognition target database, and the similarity between all vocabularies in the recognition target database and the input speech is calculated as a recognition likelihood.

  In the current recognition technology, the number of vocabulary that can be recognized within the real time of interactive processing (when the search device and the user interact, the time when the user does not feel stress or unnaturalness) is limited. If the vocabulary number exceeding the time is set as the recognition target, the processing time becomes long and the user is put on standby. On the other hand, if the number of vocabulary to be recognized is large, and if the recognition target is composed of very similar vocabularies, the recognition accuracy will be reduced so that even human dialogue may be mistaken or missed. Inevitable.

  Furthermore, the recognition accuracy greatly depends on the speaker and the speech environment, and there is no guarantee that 100% accuracy is always obtained depending on the influence of ambient noise and the like.

Therefore, in a conventional search device that uses voice recognition technology, when it is determined whether or not the candidate of the result of voice recognition of the input voice by the user is correctly recognized, the user is confirmed whether the candidate is correct or incorrect. The candidate presentation is repeated until correct answers can be confirmed in descending order of recognition likelihood.

  In a search device that uses a speech recognition device to search and determine a search term intended by a user from a predetermined database, the number of search terms constituting the database can be processed within the real time of interactive processing. If the number is greater than the number, the user is put on standby during the recognition process, and after this standby, the candidate output by the recognition device is presented to the user, and correctness is confirmed to determine whether or not the recognition has been correctly performed. .

  There is a problem for the user that waiting after an utterance and being forced to confirm the correctness after this waiting leads to unnaturalness of the dialogue and causes great stress.

  On the other hand, in an operator-compatible system, even if a mistake is made or missed, the intention can be communicated to the operator in the course of a natural conversation. Does not occur.

  In a search device that searches for a search term input by voice, in order to acquire user satisfaction, it is possible to respond in real time like an operator, and moreover, it is possible to accurately grasp the intention of the user, above all, It is necessary to determine the user intention in the course of natural dialogue with the user.

In the present invention, when a compound noun exceeding the number of search words that can be processed in the real time of interactive processing is registered in the search database, and a compound noun having many similar candidates is registered in the search database, It is an object of the present invention to provide a search apparatus, a search method, and a database for a compound noun inputted by voice that can efficiently search for a search word inputted by a user in a short time.

The present invention provides a search information database in which search information, which is character information to be searched, is stored, and beats extracted from the search information stored in the search information database. The first beat, which is the beat constituting the head of the information, is stored in the order of frequency used in the search information database, and the beat database that is stored in the order of frequency of use for the beat following the first beat, and the user When the search information is inputted by voice, the recognition processing means for preferentially processing the first beat in the beat database from the frequently used beats in the search information database, and the recognition result as a result of recognizing the first beat. The first beat whose first beat recognition score exceeds the first threshold is determined to be a leading first beat, and is a beat following the leading first beat, The recognition process is preferentially recognized from the high beats, the leading second beat is extracted, and the recognition process repeating means for repeating the recognition process until the last beat constituting the search information, and the recognition process for all beats are completed. A total score calculating means for calculating the total score of the first beat score and the second beat score for each search information output at the time, and if the total score exceeds the second threshold, the actual search word candidate list It is a voice interactive information search device having dialog means for executing a dialog with the user and confirming the search information by displaying the search word candidate in the predetermined display.

According to the present invention, it is possible to extract recognition results in the order of search words having a high existence probability without causing the user to wait.

  The best mode for carrying out the invention is the following examples.

  FIG. 1 is a block diagram showing a search apparatus 1 for compound nouns inputted by speech according to the first embodiment of the present invention.

  The compound noun search device 1 inputted with speech includes a speech input unit 2, speech recognition unit 3, speech recognition software 3S, speech recognition result output unit 4, real search word candidate list creation unit 5, dialogue A control unit 6, an audio output unit 7, audio output software 7 </ b> S, and a system database 8 are included.

  In the compound noun search device 1 inputted by voice, the voice of the user P inputted via the voice input unit 2 is sent to the voice recognition unit 3, and the voice recognition unit 3 performs voice recognition processing on the input voice. In addition, the system database 8 is used. In addition, the voice recognition unit 3 performs a recognition process on the voice input by the user P using the voice recognition software 3S.

  The system database 8 includes a search database 81, a search auxiliary database 82, and a YES / NO database 83.

  The search database 81 is a database in which compound nouns composed of a plurality of single nouns are registered as search words, and the search words are registered by being divided into single nouns.

  In the search auxiliary database 82, single nouns constituting all the search words registered in the search database 81 are registered in the order of the frequency, and the single nouns registered in the order of the frequency are The database is divided into a predetermined number (500 in the search device 1) in order from the highest to form a plurality of single noun sets.

  The YES / NO database 83 is a database that recognizes the contents (for example, yes / no, yes / no) that the user P has responded to.

  The voice recognition software 3S selects the search database 81 or the search auxiliary database 82 from the system database 8 in accordance with the processing scene of the search device 1.

  When the search term is inputted by voice, the search auxiliary database 82 is referred to, and when the response to the correctness confirmation to the user P is recognized, the YES / NO database 83 is referred to.

  The voice recognition unit 3 uses the voice recognition software 3S during the voice recognition process, and the voice output unit 7 uses the voice output software 7S during the voice output.

  Next, the search device 1 will be described more specifically.

  In the following description, a case will be described in which a search device 1 for compound nouns input by voice is provided in an input interface portion of a service for determining a corporate name such as a company name.

  There are 22 million corporate names listed in phone books across Japan, and it takes a considerable amount of time to identify one corporate name out of 22 million. Therefore, with current speech recognition technology, It is impossible to recognize and process the corporate name within the real time of dialogue processing (when the search device and the user P interact, the time when the user P does not feel stress or unnaturalness) In this case, since there are many very similar data, the recognition accuracy is very low.

  FIG. 2 is a diagram showing a specific example in which the corporate name in the search database 81 is composed of compound nouns, and these compound nouns are registered by dividing each single noun constituting the compound nouns.

  The frequency of each noun constituting the corporate name registered in the search database 81 is examined, and single nouns are recorded in order of frequency, and the search auxiliary database 82 is created. That is, the search auxiliary database 82 is a database in which the frequency of each single noun constituting the corporate name registered in the search database 81 is counted and registered in order of frequency.

  The total number of single nouns constituting the 22 million corporate names is 6.5 million.

  FIG. 3 is a diagram showing a specific example of the search assistance database 82 in the search device 1.

  The single nouns in the auxiliary search database 82 are single nouns constituting the compound nouns registered in the search database 812, arranged in order of frequency of use, and picked out 500 in descending order of frequency of use. Is a single noun set group.

  The set group including the single noun with the highest use frequency is set as the first single noun set group G1 (single noun set group that is a target to be recognized with priority over other single noun set groups), and then the use frequency A single noun set group including single nouns having a high value is referred to as a second single noun set group G2, and the third single noun set group G3, the fourth single noun set group G4,... And

  Assuming that the number of nouns constituting one set group is n, in the search device 1, n = 500, and this number is determined as follows. That is, a single noun that can be processed by the speech recognition software 3S within the real time of interactive processing (when the search device and the user P interact, it is a time during which the user P does not feel stress or unnaturalness). When the number is T and the average number of single nouns of all search words registered in the search database 81 is M, n = T / M.

  Specifically, the number T of single nouns that can be processed by the speech recognition software 3S within the real time of the interactive processing is 1500, and the average number of single nouns M of all search words registered in the search database 81 is 3. In this case, n = T / M = 1500/3 = 500.

  The number of single nouns constituting each single noun set group G1, G2, G3, G4,... May be a number other than 500, but can be processed within the real time of interactive processing as described above. It is determined by the number T of single nouns (the performance of the speech recognition software 3S) and the average number M of single nouns included in the search terms registered in the search database 81.

  FIG. 4 is a flowchart specifically showing a processing process for determining a search word in the search device 1.

  As this premise, it is assumed that a search database 81 and a search auxiliary database 82 have been created.

  First, the user P is requested to input the name of a corporation (compound noun) to be searched in a form divided into single nouns (S0). A function k indicating the rank of the single noun set group is set to 1.

  Consider the case where the compound noun “Yokosuka / Citizen / Hospital” is input as a search term by voice. First, for each of the nouns “Yokosuka”, “Citizen”, and “Hospital” inputted by voice, the voice recognition unit 3 performs a recognition process using the first single noun set group G1 (S1). That is, the speech recognition unit 3 performs recognition processing within the range of the first single noun set group G1. Then, the speech recognition result output unit 4 outputs the recognition result.

  Among the single nouns “Yokosuka”, “Citizens”, and “Hospitals” inputted by voice, the single nouns “Yokosuka” and “Hospital” are included in the first single noun set group G1, but “Citizens” As shown in FIG. 3, the frequency rank is 790, so it is included in the first single noun set group G1 and included in the second single noun set group G2.

  FIG. 5 is a diagram showing a specific example of a recognition result list in which the recognition result for each single noun constituting the search word (compound noun) input by speech and the recognition likelihood correspond to each other in the search device 1. is there.

  When one single noun is input by speech, the speech recognition result output unit 4 outputs a plurality of constituent noun candidates for the input single noun and the recognition likelihood for each of the plurality of constituent noun candidates. The recognition result list is created by associating the outputted constituent noun candidates with the recognition likelihood (S1). The recognition result list created in this way is shown in FIG.

  For example, when the simple noun “Yokosuka” is input by speech, the speech recognition unit 3 converts the input speech to “Yokosuka”, “Yokohama”, “Sukahama”, “Yokosoko”, “ ……, “Yokosuka” has a recognition likelihood of 97, “Yokohama” has a recognition likelihood of 90, “Sukahama” has a recognition likelihood of 89, and “Yokosuka” has a recognition likelihood of “Yokosuka”. The degree is 80.

  Note that the recognition likelihood for the constituent noun candidates is determined in advance when the search database 81 is created.

  Here, a constituent noun candidate having a recognition likelihood equal to or greater than a predetermined threshold is set as a dominant constituent noun candidate. In the search device 1, the predetermined threshold is set to 80. That is, a constituent noun candidate having a recognition likelihood of 80 or more is a dominant constituent noun candidate.

  And from each recognition result shown in FIG. 5, the single noun whose recognition likelihood is 80 or more is selected as a leading constituent noun candidate (S2).

  As shown in FIG. 5, with respect to the single noun “Yokosuka” input as speech, four candidates “Yokosuka”, “Yokohama”, “Sukahama”, and “Yokosana” are selected as probable constituent noun candidates and input as speech For the single noun “Citizen”, one candidate of “City” was selected as a probable constituent noun candidate, and for the single noun “Hospital” input by voice, two candidates “Hospital” and “Beauty Salon” are prominent Selected as a noun candidate. Using these influential constituent noun candidates, an actual search word candidate list is created (S3). The actual search word candidate list is a list of search word candidates that are presented to the user P and requested to be checked among the search words actually existing in the search database 81.

  By the way, since the frequency ranking of the single noun “citizen” in the auxiliary database 82 is 790 as shown in FIG. 2, the first single noun set group G1 does not include the single noun “citizen”. . Therefore, the recognition result for “citizen” shown in FIG. 5 does not include the single noun “citizen”.

  FIG. 6 is a diagram showing how the search device 1 creates a real search word candidate list.

  When a single noun with a recognition likelihood of 80 or more is combined among the recognition results shown in FIG. 5, the combination of single nouns registered in the search database 81 is “Yokosuka / City / Hospital”. ”And“ Yokohama / City / Hospital ”, and these two combinations are selected as actual search word candidates that are search word candidates existing in the search database 81. When the real search word candidates selected in this way are listed, the real search word candidate list shown at the bottom of FIG. 6 is created (S3).

  In the search device 1, the search word recognition likelihood is obtained by adding the recognition likelihood of each single noun as shown in FIG. 6. When calculating the search word recognition likelihood, the recognition likelihood of each single noun may be multiplied.

  Then, it is determined whether or not there exists a candidate in the real search word candidate list that satisfies a condition capable of determining the search word only by the confirmation process by the user P (S21).

  Here, if a threshold value of the search word recognition likelihood is determined in advance and the search word recognition likelihood of the predetermined search word candidate exceeds the threshold value of the search word recognition likelihood, the search word candidate is determined to be the user P It is determined that the condition for determining the search term is satisfied only by the confirmation process by the above. In the search device 1, it is assumed that the threshold value of the search word recognition likelihood is 270.

  Then, the search word candidate is presented to the user P on a display or the like, and confirmation is requested (S22). If the user P responds YES (S23), the search process is terminated.

  Since none of the search word candidates in the real search word candidate list shown in FIG. 6 exceeds the threshold 270 of the search word recognition likelihood, the search word candidates in the real search word candidate list shown in FIG. In all cases, the search term cannot be determined only by the confirmation process by the user P.

  In preparation for the case where the search term cannot be determined in this way, processing such as recognition of single nouns and calculation of recognition likelihood is performed in advance within the range of the second single noun set group G2. That is, in the range of the second single noun set group G2, a single noun is presented in parallel with the process of presenting search word candidates to the user P on a display or the like and requesting confirmation (S22) and the response by the user P (S23). Recognition, likelihood calculation (S12), re-selection of potential constituent noun candidates (S13), and update of the actual search word candidate list (S14).

  That is, first, the function k indicating the rank of the single noun set group is incremented by 1 (S11), k = 2, and the second most frequently used next to the first single noun set group G1 for which the recognition process is completed. Recognize in the range of the single noun set group G2 (single noun set group composed of single nouns having frequency ranks from 500 to 1,000), and use this recognition result to Extraction is attempted (S12, S13, S14).

  In the flowchart shown in FIG. 4, after step S <b> 3, the route to step S <b> 11 is indicated by a broken line separately from the route to step 21. This means that the routine of steps S21 to S23 and the routine of steps S11 to S15 are processed in parallel. That is, while the dialogue process (S21 to S23) with the user P is being executed for the first single noun set group G1, re-selection and real existence of influential constituent noun candidates for the second single noun set group G2. The search word candidate list is updated (S11 to S15).

  After that, if necessary, while the dialogue processing (S21 to S23) with the user P is being executed for the second single noun set group G2, the third single noun set group G3 is influential. Reselection of constituent noun candidates, update of the actual search word candidate list, etc. (S11 to S15) are executed, and further, dialogue processing with the user P (S21 to S23) is executed for the third single noun set group G3. In the meantime, for the fourth single noun set group G4, re-selection of probable constituent noun candidates, update of the actual search word candidate list, etc. (S11 to S15) are executed. These parallel processes are further repeated as necessary.

  FIG. 7 is a diagram illustrating a specific example of the updated recognition result list in the search device 1.

  FIG. 7 shows a list in which recognition results recognized in the range of the second single noun set group G2 are added, and influential constituent noun candidates are updated.

  As shown in FIG. 2, since the frequency ranking of the single noun “citizen” in the auxiliary database 82 is 790, the single noun “citizen” is included in the second single noun set group G2. In the updated recognition result list shown in FIG. 7, “citizen” is included as a leading constituent noun candidate in the recognition result for the single noun “citizen” inputted by voice.

  FIG. 8 is a diagram illustrating the actual search word candidate list obtained as a result of executing the process of creating the actual search word candidate list in the search device 1.

  As a result of executing the process of creating the real search word candidate list, the real search word candidate list shown in FIG. 8 is obtained, and the search word candidates having the search word recognition likelihood exceeding the threshold value 270 are “Yokosuka / “Citizen / Hospital” and “Yokohama / Citizen / Hospital”. Of these, “Yokosuka / Citizen / Hospital” is presented to the user P and the correctness is confirmed (S22).

  In a right / wrong confirmation, if a response (YES) indicating affirmation is obtained from the user P (S23), the process is terminated as if the search word has been confirmed.

  By the way, when the number of vocabulary for speech recognition is large and there are many similar vocabularies, the conventional speech recognition technology takes a long recognition processing time. It is very difficult to search within a certain time without giving stress to the user P, and the accuracy of the recognition apparatus is inevitably lowered.

  Therefore, the search device 1 instructs the user P to input the search term by limiting the search target (search term) to only a compound noun composed of a single noun sequence and separating the single noun one by one. For each single noun input by the user P, recognition processing is performed using the search auxiliary database 82 that holds the single nouns constituting the search word in order of use frequency. This is a feature of the search device 1.

  Although it is expected that the total number of single nouns registered in the search auxiliary database 82 is smaller than the total number of search terms registered in the search database 81, the total number of single nouns registered in the search auxiliary database 82 Is often a number that cannot be processed within the real time of interactive processing. In this way, the total number of single nouns registered in the search auxiliary database 82 is a number that cannot be processed within the interactive processing real time, and the single nouns registered in the search auxiliary database 82 are processed at a time. If it tries to do so, the user P will have to wait.

  Therefore, the search device 1 arranges the single nouns in the search auxiliary database 82 in order of frequency of use, and divides the single nouns into a predetermined number that can be processed within the real time of the interactive processing, thereby providing a plurality of single noun sets. Are recognized in order of frequently used single nouns, and the recognition processing for the remaining single noun sets is performed during the dialogue for selecting the search result candidate recognition results for the single noun sets. As a result, as soon as the first search word candidate selection dialogue is completed for one search word, the second search word candidate selection dialogue is performed, and if necessary, the third, fourth,... Search word candidates. Since the selection dialogue is continued, there is no time for the user P to wait.

  In other words, since the user P cannot see the processing status inside the search device 1, it looks as if the search terms registered in the search database 81 are collectively recognized. .

  Moreover, according to the search device 1, from the recognition result for each single noun, a powerful constituent noun candidate is selected, and only the search word candidates actually existing in the search database 81 are selected from all the combinations of the selected leading constituent noun candidates. And a search word candidate that can be determined to satisfy a condition that can be determined only by the confirmation process by the user P is presented to the user P. Therefore, the search device 1 can extract the recognition results in the order of search words having a high existence probability.

  That is, in the search device 1, by reducing the number of single nouns to be recognized at a time, the user P is not put on standby, and the recognition nouns are processed in order from the most frequently used single nouns. The recognition results can be extracted in the order of search words having the highest existence probability.

  In other words, in the search device 1, when searching for compound nouns using speech recognition, the length of the recognition processing time resulting from the majority of recognition targets is not perceived by the user P. The user P is not made to feel the stress caused by the necessity of repeating correct / incorrect, and yet has high-precision naturalness.

  FIG. 9 is a diagram generally showing a search database 81 in the search device 1.

  The search database 81 holds a number of search words that cannot be subjected to speech recognition processing within the real time of interactive processing, and each search word is registered in a form divided into single nouns.

That is, for example, a compound noun N ₁ that is a search word is configured by single nouns n ₁ , n ₂ , and n ₃ , and the search database 81 includes single nouns n ₁ , n ₂ , and n ₃ as compound nouns N _1. , It is registered in a state of being separated for each single noun.

  FIG. 10 is a diagram generally showing the search assistance database 82.

  The search auxiliary database 82 is a database in which the frequency of use of each single noun composing all the search terms registered in the search database 81 is calculated, and single nouns are registered in a state of being arranged in the order of frequency.

  Although the number of single nouns registered in the search auxiliary database 82 is expected to be smaller than the number of search words in the search database 81, the number of single nouns is configured such that it cannot be processed within the interactive processing real time.

  In the auxiliary search database 82, single nouns are arranged (in order) in descending order of frequency, and the single nouns are divided by the number of single nouns that can be processed within a predetermined dialogue processing real time (the above-mentioned order of single nouns). And a single noun set group is formed. A single noun set group including the most frequent single noun is referred to as a first single noun set group G1.

  Further, the search device 1 can be grasped as an invention of a recording medium.

  That is, the search device 1 creates a search database that creates a search database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are divided into individual nouns and registered. Procedures and single nouns that make up all search terms registered in the search database are registered in order of their frequency, and single nouns registered in order of their frequency are In order, a search auxiliary database creation procedure for creating a search auxiliary database in which a plurality of single noun sets are formed divided into a predetermined number and a user P inputs the search word for each single noun, For each input single noun, recognition processing is performed within the range of the first single noun set group including the single noun having the highest frequency among the plurality of single noun set groups, and the recognition likelihood is made to correspond. Recognition result list creation procedure for creating a recognition result list, a recognition result in which pairs of constituent noun candidates that are the recognized nouns and recognition likelihoods for the constituent noun candidates are arranged in order of recognition likelihood A list is created for each of the single nouns input by speech, and among the constituent noun candidates described in the recognition result list, the constituent noun candidates having a recognition likelihood exceeding a predetermined threshold are influential components. Select only the search words that exist in the search database from all possible combinations of nominal constituent noun candidates and the probable constituent noun candidate selection procedure to be selected as noun candidates, and set the recognition likelihood of each influential noun candidate to a predetermined value. An actual search word candidate list creation unit that performs calculation, calculates a search word recognition likelihood, and creates a real search word candidate list arranged in descending order of the calculated search word recognition likelihood If the search word candidate in the real search word candidate list satisfies the condition that the search word can be specified only by the confirmation process by the user P, the necessary dialogue is executed with the user P. Then, it is an example of a computer-readable recording medium recording a program for causing a computer to execute a search word candidate selection dialog procedure for determining a search word.

  As the recording medium, FD, CD, DVD, HD, semiconductor memory, and the like are conceivable.

  In addition, in the input interface part in the product delivery service, telephone number search, postal code search, etc., the corporate name determination work performed by the search device 1 can be widely applied.

  FIG. 11 is a block diagram showing a compound noun search device 101 with speech input according to the second embodiment of the present invention.

  The compound noun search device 101 input by speech includes a speech input unit 102, a speech recognition unit 103, speech recognition software 103S, a speech recognition result output unit 104, an actual search word candidate list creation unit 105, a dialogue A control unit 106, an audio output unit 107, audio output software 107S, and a system database 108 are included.

  In the compound noun search device 101 input by voice, the voice of the user P input via the voice input unit 102 is sent to the voice recognition unit 103, and the voice recognition unit 103 performs voice recognition processing on the input voice. In addition, the system database 108 is used. In addition, the voice recognition unit 103 performs recognition processing on the voice input by the user P using the voice recognition software 103S.

  The system database 108 includes a search database 181, a search auxiliary database 182, and a YES / NO database 183.

  The search database 181 is a database in which compound nouns composed of a plurality of single nouns are registered as search words, and the search words are registered by being divided into single nouns.

  The search auxiliary database 182 refers to a group of single nouns written in the nth (n is an integer value) of each registered compound noun as a group of single nouns represented by the nth notation. Single noun sets are registered in descending order of frequency, and a predetermined number (500 in the second embodiment) of single nouns registered in descending order of frequency is registered. ) Is a search auxiliary database in which a plurality of sub-set groups are formed.

  The YES / NO database 183 is a database that recognizes the contents (for example, yes / no, yes / no) that the user P has responded to.

  The speech recognition software 103S selects the search database 181 or the search auxiliary database 182 from the system database 108 in accordance with the processing scene of the search device 101.

  When the search term is inputted by voice, the search auxiliary database 182 is referred to, and when the response to the correctness confirmation to the user P is recognized, the YES / NO database 183 is referred to.

  The voice recognition unit 103 uses the voice recognition software 103S during the voice recognition process, and the voice output unit 107 uses the voice output software 107S during the voice output.

  Next, the second embodiment will be described more specifically.

  In the following description, a case will be described in which a search device 101 for compound nouns input by voice is provided in the input interface portion of a service for determining a corporate name such as a company name.

  There are as many as 22 million corporate names listed in phone books across Japan, and it takes a considerable amount of time to identify one corporate name out of 22 million based on the input voice. In the current speech recognition technology, the corporate name can be recognized and processed within the real time of dialogue processing (when the search device and the user P interact, the time when the user P does not feel stress or unnaturalness). In this case, the recognition accuracy is very low because there are many very similar data.

  FIG. 12 is a diagram showing a specific example in which the corporate name in the search database 181 is composed of compound nouns in the second embodiment, and these compound nouns are registered separately for each single noun constituting the compound nouns. is there.

  The frequency of each noun constituting the corporate name registered in the search database 181 is examined, and single nouns are recorded in order of frequency, and the search auxiliary database 182 is created.

  The total number of single nouns constituting the 22 million corporate names is 6.5 million. In addition, the total number of single nouns (single nouns located in the first word in the corporate name) of nouns constituting 22 million corporate names is about 3.6 million. The total number of single nouns (single nouns located in the second word in the corporate name) is about 2.5 million, and the third single noun (single noun located in the third word in the corporate name) ) Is about 2.7 million types, and the total number of single nouns in the fourth notation (single noun located in the fourth word in the corporate name) is about 1 million types, ... (the maximum number of constituent words is 7) In the real time of dialogue processing, it is impossible to recognize each single noun with respect to one compound noun, and it is expected that the accuracy is low.

  FIG. 13 is a diagram showing a specific example of data stored in the auxiliary search database 182 in the second embodiment.

  Single nouns in the search auxiliary database 182 are single nouns constituting compound nouns registered in the search database 181, and a single noun set group is formed for each notation order in the compound nouns. By collecting only the single nouns whose notation order is the first in the compound noun, the first noun single noun set group Ga1 is formed, and in the first noun single noun set group Ga1, the single nouns are used in descending order of frequency of use. In addition, these registered single nouns are grouped into 500 single nouns in order from the highest frequency to form a plurality of sub-set groups, and the first sub-set in descending order of frequency. A group Ga1-1, a second sub-set group Ga1-2,..., An m-th sub-set group Ga1-m are formed.

  The first subset group Ga1-1 including the single noun having the highest use frequency is a target to be recognized with priority over the other subset groups Ga1-2 to Ga1-m.

  Further, only the single nouns with the second notation order in the compound nouns are collected to form the second notation single noun set group Ga2, and the second noun single noun set group Ga2 is used in descending order of frequency of use. Single nouns are registered, and these registered single nouns are grouped in units of 500 single nouns in descending order of their frequency, and a plurality of sub-set groups are formed. One sub-set group Ga2-1, second sub-set group Ga2-2,..., M-th sub-set group Ga2-m are formed.

  The first subset group Ga2-1 including the single noun having the highest use frequency is a target to be recognized with priority over the other subset groups Ga2-2 to Ga2-m.

  Hereinafter, in the same manner as described above, only the single nouns in which the order of the notation in the compound noun is n-th are collected to form the noun single-noun set group Gan. Single nouns are registered in descending order of frequency of use, and the registered single nouns are grouped in units of 500 single nouns in order from the highest frequency to form a plurality of sub-set groups. The first sub-set group Gan1, the second sub-set group Gan2,..., The n-th sub-set group Gan-m are formed in descending order.

  The first subset group Gan-1 including the single noun having the highest use frequency is a target to be recognized with priority over the other subset groups Gan-2 to Gn-m.

  By the way, when the number of nouns constituting one sub-set group is N, in the second embodiment, N = 500, and this number is determined as follows. That is, a single noun that can be processed by the speech recognition software 103S within the real time of dialogue processing (when the search device and the user P interact, it is a time during which the user P does not feel stress or unnaturalness). When the number is T and the average number of single nouns of all search words registered in the search database 181 is M, N = T / M.

  Specifically, it is assumed that the number T of single nouns that can be processed by the speech recognition software 103S within the real time of dialogue processing is 1500, and the average number of single nouns M of all search words registered in the search database 181 is 3. N = T / M = 1500/3 = 500.

  The number N of single nouns constituting each of the sub-groups Ga1-1 to Gan-m may be other than 500, but as described above, the single nouns that can be processed within the interactive processing real time. It is determined by the number T (performance of the speech recognition software 103S) and the average number M of single nouns included in the search terms registered in the search database 181.

  FIG. 14 is a flowchart specifically showing a processing process for determining a search term in the second embodiment.

  As this premise, it is assumed that a search database 181 and a search auxiliary database 182 have been created.

  First, the user P is requested to input the name of the corporation (compound noun) to be searched in a form divided into single nouns (S100). Then, the function k of the sub-group in each single-noun group related to the compound noun that is the search word (function indicating the rank of the sub-group) is set to 1. That is, if the function k = 1 of the subset group, each single noun is recognized using the subset groups Ga1-1, Ga2-1, Ga3-1,... Among the plurality of subset groups. Is done.

  Consider the case where the compound noun “Yokohama / Grand / Hotel” is input as a search term by voice.

  First, for each of the nouns “Yokohama”, “Grand”, and “Hotel” inputted by speech, the first sub-noun group Ga1-1 and the second-noun single noun set in the first-noun single noun set group Ga1 The speech recognition unit 103 performs a recognition process using the first sub-set group Ga2-1 in the group Ga2 and the first sub-set group Ga3-1 in the third-named single noun set group Ga3 (S101). That is, the single noun “Yokohama” input by speech is recognized within the range of the first sub-set group Ga1-1, and the range of the first sub-set group Ga2-1 is determined for the single noun “Grand” input by speech. The single noun “hotel” input by voice recognition is recognized and processed within the range of the first sub-set group Ga 3-1. Then, the speech recognition result output unit 104 outputs the recognition result.

  Of the single nouns “Yokohama”, “Grand”, and “Hotel” input by voice, the single noun “Yokohama” is included in the first sub-group G 1-1 as shown in FIG. It is assumed that “hotel” is also included in the first sub-group G <b> 3-1. However, as shown in FIG. 13, the frequency rank of the single noun “Grand” is the frequency rank 951, so it is not included in the first subset group Ga2-1 and included in the second subset group Ga2-2. It is.

  FIG. 15 shows a specific example of a recognition result list in which the recognition results for the single nouns constituting the search words (compound nouns) inputted by speech are associated with the recognition likelihoods in the second embodiment. FIG.

  When one single noun is input by speech, the speech recognition result output unit 104 outputs a plurality of constituent noun candidates for the input single noun and the recognition likelihood for each of the plurality of constituent noun candidates. The recognition result list is created by associating the outputted constituent noun candidates with the recognition likelihood (S101). The recognition result list created in this way is shown in FIG.

  For example, when the single noun “Yokohama” is input by voice, the voice recognition unit 103 inputs the input voice as shown in FIG. 15 as “Yokohama”, “Yokosuka”, “Yokoyama”, “Yokoyama”,. The recognition likelihood of “Yokohama” is 95, the recognition likelihood of “Yokosuka” is 90, the recognition likelihood of “Yokomizo” is 81, and the recognition likelihood of “Yokoyama” is 75. is there.

  Note that the recognition likelihood for each single noun is individually determined by the voice recognition device 3S when the single noun is input by voice.

  Here, a constituent noun candidate having a recognition likelihood equal to or greater than a predetermined threshold is set as a dominant constituent noun candidate. In the second embodiment, the predetermined threshold is 80. That is, a constituent noun candidate having a recognition likelihood of 80 or more is a dominant constituent noun candidate. That the predetermined threshold is 80 is set in advance when the databases 81 and 82 are created.

  And from each recognition result shown in FIG. 15, the single noun whose recognition likelihood is 80 or more is selected as an influential constituent noun candidate (S102).

  As shown in FIG. 15, with respect to the single noun “Yokohama” inputted by speech, three candidates “Yokohama”, “Yokosuka”, and “Yokomizo” are selected as potential constituent noun candidates, and the single noun “Grand” "Is not selected as a prominent constituent noun candidate, and for the single noun" hotel "input by speech, two candidates" hotel "and" hall "are selected as prominent constituent noun candidates.

  Then, an actual search word candidate list is created using these potential constituent noun candidates (S103). The actual search word candidate list is a list of search word candidates that are presented to the user P and asked for confirmation among the search words existing in the search database 181.

  By the way, the frequency ranking of the second noun single noun “Grand” is ranked 951 in the second noun single noun set group Ga2 stored in the auxiliary database 82, as shown in FIG. The single noun “grand” is not included in the first sub-set group Ga2-1 of the single noun set group Ga2. Therefore, the single noun “ground” is not included in the recognition result for the voice-input single noun “ground” shown in FIG.

  Therefore, in the step (S103) of creating the actual search word candidate list, the actual search word candidate list is not created.

  Since the condition for specifying the search word is not satisfied only by the confirmation process (S121), the function k of the subset group is incremented by 1 (S111), and the subset group to be searched is Ga1-2, Ga2-2, Each single noun that becomes Ga3-1 and is input by speech is recognized using the new set group Ga2-2, and the recognition likelihood is calculated (S112). Then, with the newly recognized recognition result added (using the updated recognition result list), the potential constituent noun candidates are re-selected (S113) and the actual search word candidate list is updated (S114).

  FIG. 16 is a diagram illustrating a specific example of the updated recognition result list in the second embodiment.

  FIG. 16 shows a list in which recognition results recognized in the range of the second subset group Ga1-2, Ga2-2, and Ga3-2 are added, and influential constituent noun candidates are updated.

  As shown in FIG. 13, in the auxiliary database 82, the frequency ranking of the single noun “Grand” in the second-named single noun set group Ga2 is 951, so the single noun “Grand” is simply assigned to the second sub-set group Ga2-2. The noun “Grand” is included, and therefore, the updated recognition result list shown in FIG. 16 includes “Grand” as a leading constituent noun candidate in the recognition result for the single noun “Grand” inputted by speech. It is.

  FIG. 17 is a diagram showing how a real search word candidate list is created in the second embodiment.

  Recognize a new sub-set group in the speech input single noun “Grand” that could not select influential constituent noun candidates, and combine single nouns with a recognition likelihood of 80 or more among the recognition results. Among them, combinations of single nouns registered in the search database 181 are two combinations of “Yokohama / Grand / Hotel” and “Yokohama / Gold / Hall” as shown in FIG. The combination is selected as an actual search word candidate that is a search word candidate existing in the search database 181. When a real search word candidate list is created for the real search word candidates selected in this way, the real search word candidate list shown at the bottom of FIG. 17 is updated (S114).

  In the second embodiment, the search word recognition likelihood is obtained by adding the recognition likelihood of each single noun as shown in FIG. When calculating the search word recognition likelihood, the recognition likelihood of each single noun may be multiplied.

  Then, it is determined whether or not there exists a candidate in the real search word candidate list that satisfies a condition that can determine the search word only by the confirmation process by the user P (S121).

  Here, if a threshold value of the search word recognition likelihood is determined in advance and the search word recognition likelihood of the predetermined search word candidate exceeds the threshold value of the search word recognition likelihood, the search word candidate is determined to be the user P It is determined that the condition for determining the search term is satisfied only by the confirmation process by the above. In the second embodiment, it is assumed that the threshold for the search word recognition likelihood is 280.

  Then, the search word candidate is presented to the user P on a display or the like, confirmation is requested (S122), and if the user P responds YES (S123), the search process is terminated.

  Since one candidate among the search word candidates in the real search word candidate list shown in FIG. 17 satisfies the condition that the search word recognition likelihood threshold value is 280 or more, the candidate in the real search word candidate list shown in FIG. Search word candidates can be determined by only the confirmation process by the user P.

  If none of the search word candidates in the real search word candidate list satisfies the condition that the search word recognition likelihood threshold value is 280 or more, all the search word candidates in the real search word candidate list are The search term cannot be determined only by the confirmation process by the user P.

  In preparation for the case where the search term cannot be determined in this manner, processing such as recognition of a single noun and calculation of recognition likelihood in the range of the third subset group Ga1-3, G2-3, G3-3, Perform in advance. In other words, in parallel with the process of presenting the search word candidate to the user P on the display and requesting confirmation (S122) and the response by the user P (S123), the third subset group Ga1-3, Ga2-3, In the range of Ga3-3, recognition of single nouns, calculation of recognition likelihood (S112), re-selection of potential constituent noun candidates (S113), and update of the actual search word candidate list (S114) are executed.

  That is, the function k indicating the order of the subset group is further incremented by 1 (S111), k = 3, and the third subset group Ga1-3 having the second highest usage frequency after the second subset group after the recognition process is completed. , Ga2-3, Ga3-3 (recognized in the range of sub-groups Ga1-3, Ga2-3, Ga3-3 composed of single nouns with frequency ranks of 1001 to 1,500, Using the recognition result, reextraction of probable constituent noun candidates is attempted (S112, S113, S114).

  In the flowchart shown in FIG. 14, after step S <b> 3, the route to step S <b> 1111 is indicated by a broken line separately from the route to step 21. This means that the routine of steps S1 21 to S123 and the routine of steps S111 to S115 are processed in parallel.

  That is, while the dialogue processing (S121 to S123) with the user P is being performed on the first sub-groups Ga1-1, Ga2-1, Ga3-1, the second sub-groups Ga1-2, Ga2- For 2, Ga3-2, re-selection of possible constituent noun candidates, update of the actual search word candidate list, and the like (S111 to S115) are executed.

  After that, if necessary, the third sub-group G1-2, Ga2-2, Ga3-2, while the dialog processing (S121 to S123) with the user P is being executed, For the set groups Ga1-3, Ga2-3, Ga3-3, re-selection of influential constituent noun candidates, update of the actual search word candidate list, etc. (S111 to S115) are executed, and further, the third sub-set group Ga1-3 , Ga2-3, Ga3-3, while the dialogue process (S121-S123) with the user P is being executed, the prominent constituent nouns for the fourth subset group Ga1-4, Ga2-4, Ga3-4 Re-selection of candidates, update of the actual search word candidate list, etc. (S111 to S115) are executed. These parallel processes are further repeated as necessary.

  By the way, when the number of vocabulary for speech recognition is large and there are many similar vocabularies, the conventional speech recognition technology takes a long recognition processing time. It is very difficult to search within a certain time without giving stress to the user P, and the accuracy of the recognition apparatus is inevitably lowered.

  Therefore, in the second embodiment, the search target (search term) is limited to only a compound noun consisting of a single noun sequence, and the user P is requested to input the search term by separating the single noun one by one. For each single noun entered and entered by the user P, a single noun set group is formed for each position described in the compound noun, and the single noun set group is registered in descending order of frequency. A database having a plurality of single nouns registered in order from the highest frequency and a search assisting database in which a plurality of sub-set groups are formed in order from the highest frequency in a predetermined number. Is used to search for compound nouns input by the user.

  As described above, a single noun set group is formed for each single noun according to the position of the single noun in the compound noun, and this single noun set group is divided into a plurality of sub-set groups, of which the frequency of use The feature of the second embodiment is that the single noun is recognized because it contains a high noun.

  Although it is expected that the total number of single nouns registered in the search auxiliary database 182 is smaller than the total number of search terms registered in the search database 181, the total number of single nouns registered in the search auxiliary database 182 is expected. Is often a number that cannot be processed within the real time of interactive processing. In this way, the total number of single nouns registered in the search auxiliary database 182 is a number that cannot be processed within the interactive processing real time, and the single nouns registered in the search auxiliary database 182 are processed at a time. If it tries to do so, the user P will have to wait.

  Therefore, in the second embodiment, the single nouns in the search auxiliary database 182 are arranged in the order of use frequency, and this is divided into a predetermined number that can be processed within the interactive processing real time, and a plurality of sub-sets are obtained. A group is formed, recognition processing is performed in the order of frequently used single nouns, and recognition processing for the remaining subset groups is performed at the time when the recognition result for each subset group is dialogued for selecting search word candidates. As a result, as soon as the first search word candidate selection dialogue is completed for one search word, the second search word candidate selection dialogue is performed, and if necessary, the third, fourth,... Search word candidates. Since the selection dialogue continues, there is no time to wait for the user P.

  That is, since the user P cannot see the processing status inside the search device 101, it looks as if the search terms registered in the search database 181 are being collectively recognized.

  Further, according to the second embodiment, a powerful constituent noun candidate is selected from the recognition result for each single noun, and a search word that exists in the search database 181 is selected from all the combinations of the selected leading constituent noun candidates. Only candidates are selected, and search word candidates that can be determined to satisfy conditions that can be determined only by the confirmation process by the user P are presented to the user P. Therefore, in the second embodiment, recognition results can be extracted in the order of search words having a high existence probability.

  In other words, in the second embodiment, by reducing the number of single nouns to be recognized at a time, the user P is not put on standby, and the recognition processes are performed in order from the most frequently used single noun. Thus, recognition results can be extracted in the order of search terms having a high existence probability.

  In other words, in the second embodiment, when searching for compound nouns using speech recognition, the length of the recognition processing time resulting from the majority of recognition targets is not perceived by the user P. The user P does not feel the stress caused by the necessity of repeating correct / incorrect confirmation through recognition, and has high-precision naturalness.

  By the way, “Grand”, which is the second single noun of the compound noun “Yokohama / Grand / Hotel” (the second single noun in the compound noun), arranges all the single nouns constituting all search words in order of frequency. In this case, since it is located at the 2,450th position, when all the single nouns stored in the search database 181 are arranged in the order of their use frequency and recognition processing is performed for each 500 nouns, the fifth 500 words from the top It will appear for the first time in the group. In this case, at least the recognition process and the actual search word list creation process are repeated five times. In the process, if there is a real search word candidate whose search word recognition likelihood exceeds the specified threshold, to the user Is presented, and the user P feels uncomfortable. However, in the second embodiment, “Grand” which is the second noun of “Yokohama / Grand / Hotel” (the second single noun in the compound noun) is the single noun set group Ga2 of the second notation. Since the frequency ranking is 951, if the recognition process and the actual search word list creation process are executed twice, the search word can be confirmed, and the user P does not feel uncomfortable.

  Further, the second embodiment can be grasped as an invention of a recording medium.

  That is, the second embodiment creates a search database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are registered by being divided into single nouns. A group of simple nouns described in the database creation procedure and the nth (n is an integer value) of each registered compound noun is referred to as an nth single noun set group. Noun set groups are registered in descending order of frequency, and single nouns registered in order of increasing frequency are grouped into a predetermined number in order from the highest frequency. A search auxiliary database creation procedure for creating a formed search auxiliary database, and when a user inputs a compound word that is the search word for each single noun, the n noun of the compound noun Th Recognized by the single noun set group described above, and the recognition process is performed within the range of the first sub-set group including the single noun set group with the highest frequency among the above-mentioned single noun set groups. The recognition result list creation procedure for creating the recognition result list, the constituent noun candidates that are the recognized single nouns, and the recognition likelihood for the constituent noun candidates are arranged in the order of recognition likelihood. A recognition result list for each of the single nouns input by speech, and among the constituent noun candidates described in the recognition result list, the constituent noun candidates having a recognition likelihood exceeding a predetermined threshold are selected. , Select only the search words that exist in the search database from all possible combinations of the powerful constituent noun candidates and the probable constituent noun candidate selection procedure to be selected as the dominant constituent noun candidates, Performing a predetermined operation on the cognitive likelihood, calculating a search word recognition likelihood, and creating a real search word candidate list arranged in descending order of the calculated search word recognition likelihood; and If the search word candidate in the real search word candidate list satisfies the condition that the search word can be specified only by the confirmation process by the user, a necessary dialogue is executed with the user, and the search word 5 is an example of a computer-readable recording medium that records a program that causes a computer to execute a search word candidate selection dialog procedure for determining the search term candidate.

  Here, FD, CD, DVD, HD, semiconductor memory, etc. can be considered as the recording medium.

  It should be noted that the corporate name determination work performed in the second embodiment can be widely applied to the input interface portion in the product delivery service, telephone number search, postal code search, and the like.

  FIG. 18 is a block diagram showing a search device 201 for compound nouns inputted by speech according to the third embodiment of the present invention.

  The compound noun search device 201 input by voice includes a voice input unit 202, a voice recognition unit 203, a voice recognition device 203S that uses voice recognition software, a voice recognition result output unit 204, and a voice recognition result list creation. Unit 204a, actual search word candidate list creation unit 205, dialogue control unit 206, voice output unit 207, voice output device 207S using voice output software, and system database 208.

  In the compound noun search device 201 input by voice, the voice of the user 210 input via the voice input unit 202 is sent to the voice recognition unit 203, and the voice recognition unit 203 performs voice recognition processing on the input voice. In addition, the system database 208 is used. In addition, the voice recognition unit 203 performs recognition processing on the voice input by the user 210 using the voice recognition software in the voice recognition device 203S.

  The system database 208 includes a search database 281, an order search auxiliary database 282, a reverse order search auxiliary database 283, and a YES / NO database 284.

  The search database 281 is a database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are registered by being divided into single nouns.

  In the order search auxiliary database 282, a single noun set located in the nth word, which is a set of single nouns located in the nth (n is an integer value) of each compound noun registered in the search database 281 is registered. A single noun belonging to the single noun set located in the nth word is divided in order of frequently used single nouns and every predetermined number (for example, 500), so that a plurality of nth single nouns are obtained. Subsets are constructed.

  The reverse search auxiliary database 283 includes a single noun set located in the reverse m word, which is a set of single nouns located in the reverse m word (m is an integer value) of each compound noun registered in the search database 281. A single noun belonging to a set of single nouns located in the reverse m-th word in the registered database is divided into a plurality of predetermined nouns in the order of frequently used single nouns. A single noun subset is constructed.

  The YES / NO database 284 is a database for recognizing the contents (for example, yes / no, yes / no) that the user 210 has responded to.

  The speech recognition software used in the speech recognition device 203S selects the search database 281, the order search auxiliary database 282, or the reverse order search auxiliary database 283 from the system database 208 in accordance with the processing scene of the search device 201. is there.

  When the search term is inputted by voice, the order search auxiliary database 282 or the reverse order search auxiliary database 283 is referred to, and when the response to the correctness confirmation to the user 210 is recognized, the YES / NO database 284 is referred to. .

  The speech recognition unit 203 uses the speech recognition software 203S when performing speech recognition processing, and the speech output unit 207 uses the speech output software 7S when outputting speech.

  In the following description, a case will be described in which a speech noun search device 201 is provided in an input interface portion of a service for determining a corporate name such as a company name.

  By the way, there are as many as 22 million corporate names listed in Japan's telephone directory, and it takes a considerable amount of time to identify one corporate name out of 22 million. Then, it is impossible to recognize and process the corporate name within the real time of dialogue processing (when the search device 201 and the user 210 interact, the time when the user 210 does not feel stress or unnaturalness) In addition, in this case, since there are many very similar data, the recognition accuracy is very low.

  FIG. 19 is a diagram showing a specific example in which the corporate name in the search database 281 is composed of compound nouns in the compound noun search apparatus 201, and the compound nouns are registered separately for each single noun constituting the compound nouns. It is.

  The frequency of each noun constituting the corporate name registered in the search database 281 is checked, single nouns are recorded in order of frequency, and an order search auxiliary database 282 is created. That is, the order search auxiliary database 282 is a database in which the frequency of each single noun constituting the corporate name registered in the search database 281 is counted and registered in order of frequency.

  The total number of single nouns constituting the 22 million corporate names is 6.5 million.

  When the user 210 divides a search word into single nouns and inputs a voice, the voice recognition unit 203 inputs a single noun (first single noun from the beginning) located in the first word in the compound noun input by voice, A single noun located in the second word (the first single noun from the beginning),. The first single noun), the single noun located in the reverse second word (the second single noun from the end), and so on.

  The voice recognition result output unit 204 is a part that sends the recognition result to the voice recognition result list creation unit 204a.

  The speech recognition result list creation unit 204a calculates, for each single noun, the speech recognition device 203S for the recognition result when the order search auxiliary database 282 is used and the recognition result when the reverse order search auxiliary database 283 is used. This is a part that merges using the recognition likelihood, creates a recognition result list, and sends it to the real search word candidate list creation unit 205.

  At this time, if there is a single noun for which recognition processing has not been completed yet in the order search auxiliary database 282 and reverse order search auxiliary database 283, the speech recognition unit 203 uses the single noun with the second highest frequency of use. The speech recognition result output unit 204 outputs a recognition result, sends it to the speech recognition result list creation unit 204a, creates a recognition result list, and actually exists. The process of sending to the search word candidate list creation unit 205 is repeated.

  The real search word candidate list creation unit 205 selects a potential constituent noun candidate based on the recognition result, creates all combinations of the potential constituent noun candidates, and extracts a real search term while referring to the search database 281. The actual search word candidate list is created and sent to the dialogue control unit 206.

  The dialogue control unit 206 refers to the actual search word candidate list, and confirms if the search word candidate in the actual search word candidate list satisfies a predetermined condition that allows the search word to be specified only by the confirmation process with the user 210. A processing guidance output command is sent to the audio output unit 207.

  If the search word candidate in the actual search word candidate list does not satisfy the condition for specifying the search word only by the confirmation process with the user 210, the recognition process is performed for the set of single nouns that are used the second most frequently. The update actual search word candidate list is used by using the powerful constituent noun candidates from which the processing results are obtained, and the execution of the search word candidate selection dialogue is repeated.

  When the voice output unit 207 receives the confirmation processing guidance output command, the voice output unit 207 outputs a guidance for checking the correctness of the candidate designated as the search word candidate to the user 210.

  When a response to the confirmation processing guidance is input again to the voice input unit 202, the voice recognition unit 203 refers to the YES / NO database 284, recognizes the response of the user 210, and the voice recognition result output unit 11. When the recognition result is output and a response indicating affirmation is obtained from the user 210, the dialogue control unit 206 sends a command for guiding the user 210 that the search term specification is completed to the voice output unit 207.

  Until the search terms are identified, the updating of the actual search word candidate list and the search word candidate selection dialogue are repeated, and recognition of a set of all single nouns stored in the order search auxiliary database 282 and the reverse order check plan auxiliary database 283 is performed. When the processing is completed and the actual search word candidate list cannot be updated any more, the dialogue control unit 206 presents the search word candidates to the user 210 from the top of the real search word candidate list. The guidance output command is sent to the voice output unit 207.

  The search device 201 completes the recognition process for at least one set of single nouns other than the priority recognition target while the actual search word candidate list is updated and the search word candidate selection dialogue is performed. As described above, the number of single nouns constituting the one set is defined.

  Next, a processing process for determining a search word in the compound noun search apparatus 201 will be specifically described.

  FIG. 20 is a diagram illustrating a specific example of the order search auxiliary database 282 in the compound noun search apparatus 201.

  The plurality of single nouns stored in the order search auxiliary database 282 are divided into a single noun set Gbn located at the nth word, that is, a single noun set Gb1 located at the first word and a second word. The single noun set Gb2, the single noun set Gb3 located in the third word, and so on are classified.

  In addition, a single noun belonging to a single noun set located in the nth word is divided into 500 single nouns in order of frequency of use, thereby forming a plurality of nth single noun sub-sets. The single noun set Gb1 located in the word includes a single noun sub-set Gb1-1, a single-noun sub-set Gb1-2, a single-noun sub-set Gb1-3, for every 500 single nouns in order of frequency of use. .., And the single noun set Gb2 located in the second word is divided into a single noun sub-set Gb2-1, a single noun sub-set Gb2-2, It is divided into sub-sets Gb2-3, and so on.

  In the compound noun search apparatus 201, 500 is set as the number c in the case of the division, but the number c in the case of the division is defined by the performance of the speech recognition apparatus 203S to be used.

  FIG. 21 is a diagram showing a specific example of the reverse search auxiliary database 283 in the compound noun search device 201.

  A plurality of single nouns stored in the reverse search auxiliary database 283 is divided into a single noun set Gbm located in the reverse m word. That is, the reverse single noun set H1 which is a set of single nouns (the first single noun from the last) located in the reverse-ordered first word in the compound noun input by speech and the second reverse-order word in the composite noun input by speech A reverse noun set H2 that is a set of single nouns (the second single noun from the end) and a single noun (the third single noun from the last) Noun) is a reverse-order single noun set H3, and so on.

  In addition, the single noun belonging to the single noun set located in the reverse m-th word is divided into 500 single nouns in descending order of frequency of use, so that the reverse single noun sub-set located in the multiple reverse m-words is obtained. The reverse single noun set H1 that is configured, that is, located in the reverse first word, is divided into the reverse single noun sub-set H1-1, reverse single noun sub-set H1-2, , The reverse single noun sub-set H1-3,... And the reverse single noun set H2 located in the second reverse reverse word is the reverse single noun sub-set H2- 1, a reverse single noun sub-set H2-2, a reverse single noun sub-set H2-3, and so on.

  In the compound noun search apparatus 201, 500 is set as the number c in the case of the division, but the number c in the case of the division is defined by the performance of the speech recognition apparatus 203S to be used.

  FIG. 22 is a flowchart showing a process for determining a search word in the compound noun search apparatus 201.

  First, the user 210 inputs the corporate name, which is the purpose of the search, by voice by dividing it into single nouns (S200). Consider a case where “Yokosuka / Central / Hotel” is input as a search word in the compound noun search apparatus 201.

  Here, “1” is set as a variable n indicating which word in the plural nouns the single noun constituting the compound noun is, and a single noun set Gbn located in the nth word is formed. “1” is set as the variable p of the rank (rank indicating the high usage frequency) in the single noun sub-group Gbn-p (S201).

  Then, for each of “Yokosuka”, “Central”, and “Hotel”, recognition processing is performed for the single noun subset Gb1-1, single noun subset Gb2-1, and single noun subset Gb3-1 (S202). , S203). Subsequently, for each single noun, a recognition result list in the case of performing an order search is created by associating the recognition likelihood (S204). Subsequently, single nouns with a recognition likelihood of 80 or more are selected as potential constituent noun candidates (S205).

  The “influential constituent noun candidate” is a constituent noun candidate having a recognition likelihood equal to or greater than a predetermined threshold. In the compound noun search apparatus 201, the threshold is set to 80. That is, a constituent noun candidate having a recognition likelihood of 80 or more is a dominant constituent noun candidate. Note that the recognition likelihood used when selecting a potential constituent noun candidate is predetermined when the search database 281 is created.

  FIG. 23 is a diagram illustrating a specific example of a recognition result list when an order search is performed in the compound noun search apparatus 201.

  Note that “order search” uses the order search auxiliary database 282 to recognize a single noun located at the nth word in a compound noun input by speech within the range of the single noun set Gbn located at the nth word. , The search operation.

  In addition, “reverse search” described later uses the reverse search auxiliary database 283 to convert a single noun located at the reverse m word in the compound noun inputted by speech to the single noun set Gbm located at the reverse m word. It is an operation of recognizing and searching by range.

  On the other hand, the recognition operation in the reverse search is performed in parallel with the recognition operation in the order search.

  That is, “1” is set as a variable m indicating the number of the single nouns constituting the compound noun in the reverse order in the plural nouns, and the single noun set Gbm located at the reverse m word is formed. “1” is set as the variable q of the rank (rank indicating the high usage frequency) in the single noun sub-group Gbm-q (S211).

  For each of “Yokosuka”, “Central”, and “Hotel”, recognition processing is performed for the reverse order noun sub-set H1-1, the reverse order noun sub-set H2-1, and the reverse order noun sub-set H3-1. Perform (S212, S213). Subsequently, for each single noun, a recognition result list in the case where a reverse search is performed in association with the recognition likelihood is created (S214). Then, single nouns with a recognition likelihood of 80 or more are selected as potential constituent noun candidates (S215).

  In the compound noun search apparatus 201, the recognition operation in the reverse search is performed in parallel with the recognition operation in the order search as described above. In this case, the order search (S201 to S205) is executed by one CPU. The reverse search (S211 to S215) may be executed by another CPU, or the single search may be used to execute the order search (S201 to S205) and the reverse search (S211 to S215). Also good.

  FIG. 24 is a diagram illustrating a specific example of a recognition result list when a reverse search is performed in the compound noun search apparatus 201.

  The frequency of use in the order search auxiliary database 282 is as follows. The single noun “Yokosuka” located in the first word in the compound noun input by voice is ranked 420 as shown in FIG. The noun “Central” is ranked 1250, and the single noun “Hotel” located in the third word is ranked 892. Since “Central” and “Hotel” are ranked 500 or lower, they are not included in the single noun sub-groups Gb2-1 and Gb3-1 that are priority recognition targets, and are not initially recognized.

  However, the frequency of use in the reverse search auxiliary database 283 is that the single noun “central” located in the reverse second word in the compound noun input by speech is in the ninth position as shown in FIG. Since the single noun “hotel” is ranked first, it is included in the reverse single noun sub-sets H2-1 and H1-1, is a priority recognition target, and is recognized earlier than in the case of order search.

  FIG. 25 is a diagram illustrating a specific example of a recognition result list in the case where the recognition result in the case of the order search and the recognition result in the case of the reverse search are merged in the compound noun search apparatus 201.

  When the recognition result in the case of the order search shown in FIG. 23 and the recognition result in the case of the reverse search shown in FIG. 24 are merged (S221), an integrated recognition result list is completed as shown in FIG. Select a constituent noun candidate.

  In the compound noun search apparatus 201, a single noun having a recognition likelihood of 80 or more is determined as a prominent constituent noun candidate, and as a result, as shown in FIG. Two candidates “Yokohama” and “Yokosuka” are selected as probable constituent noun candidates, and for the single noun “Central” input as speech, “Central” is selected as a prominent constituent noun candidate and input as a single noun For “Hotel”, two candidates of “Hotel” and “Hall” are selected as probable constituent noun candidates.

  FIG. 26 is a diagram showing a state immediately before the real search word candidate list is created in the compound noun search apparatus 201.

  The search device 201 uses the probable constituent noun candidates in the merged recognition result shown in FIG. 25 to perform a real search word list creation process (S222).

  The search device 201 lists actual search word candidates as shown in FIG. 26 based on all the combinations of leading constituent noun candidates and the compound nouns stored in the search database 281.

  In the compound noun search apparatus 201, “Yokosuka / Central / Hotel” (search word recognition likelihood 90 + 95 + 95 = 280) is listed as an actual search word candidate. In the compound noun search apparatus 201, the search word recognition likelihood is calculated by adding the recognition likelihood of each single noun as shown in FIG.

  In the compound noun search apparatus 201, the criteria for determining whether or not the search word can be determined only by the confirmation process with the user 210 (S223) is based on the threshold defined in advance by the compound noun search apparatus 201 and the search word recognition likelihood. A search word candidate having a search word recognition likelihood exceeding the prescribed threshold is determined as “a condition that allows the search word to be determined only by the confirmation process”.

  The actual search word candidate list is a list of search word candidates that are presented to the user 210 and requested to be checked among search words that exist in the search database 281.

  In the compound noun search apparatus 201, the prescribed threshold for the search word recognition likelihood is set to 280. In the compound noun search apparatus 201, “Yokosuka / Central / Hotel”, which is a search word candidate having a search word recognition likelihood of 280 or more, is selected from the actual search word list and presented to the user 210 (S224). If a response indicating affirmation is obtained from the user 210 in response to the correctness / incorrectness check (S225), the process is terminated assuming that the search word can be confirmed.

  By the way, when it is assumed that the number of single nouns constituting the compound noun that is a search word is 3, 4, 5,..., 8, that is, when the number of single nouns is assumed to be 8, The single nouns such as “Hotel”, “Bank”, “Center”, etc., which are likely to be used as the first single noun in reverse order of, are compound nouns composed of three single nouns. For example, it is counted as a single noun located in the third word, and if a compound noun is composed of four single nouns, it is counted as a single noun located in the fourth word, Even if 99% of the single nouns of “Bank” and “Center” are used as the first single noun in the reverse order (single noun at the end), the frequency statistics are dispersed. Therefore, as shown in FIG. 20, “hotel”, “bank”, and “center” are not subject to priority recognition with high frequency of use, and therefore, a quick search is often not performed.

  However, in the compound noun search device 201, frequency statistics are counted for each single noun position (in reverse order) counted from the end and stored in the reverse search auxiliary database 282. As shown in FIG. "Hotel", "Bank", "Center", etc. are registered as a reverse single noun set H1 located in the reverse first word that is the set of single nouns that are used most frequently from Become. In addition, “central”, “grand”, and the like are registered as the reverse-order single noun set H2 that is located in the second reverse-order word that is the set of single nouns that is used second most frequently from the end. Therefore, in the compound noun search apparatus 201, the recognition operation is executed in a short time.

  Further, in the compound noun search device 201, simultaneously with the real search word list creation process (S222), the recognition process may be started for the single noun sub-set Gbn-2 and the reverse-order single noun sub-set Hm-2. . That is, when a search term candidate is presented to the user 210 and confirmation is requested (S224), when recognition is performed in the range of the next highest single noun, the operation to be recognized from the user 210's point of view. It seems that there is no need to wait without doing anything in between, and search word candidates are presented one after another without performing recognition operation.

  If the question from the search device 201 and the response from the user 210 to this question are one turn, the compound noun search device 201 has as few dialog turns as possible with the user 210. Search term specification can be realized, thereby reducing the dialogue time and performing search processing more quickly.

  The basic search policy in the compound noun search apparatus 201 is to input a search word divided into single nouns by voice input, consider the position from the top, and perform recognition processing preferentially in descending order of frequency statistics for each single noun. By combining the recognition results for each single noun and selecting only the actual search terms, the processing time and incompleteness of accuracy of the recognition device are compensated, and the search terms are identified without stressing the user 210 It is to be.

  In addition to the basic policy described above, the major feature of the compound noun search apparatus 201 is that, in parallel with the recognition process performed for each single noun, the first word, the second word, For each third word,..., Frequency statistics are checked, and the reverse search auxiliary database 283 arranged in the order of frequency statistics is subjected to recognition processing when voice input is completed, and the recognition results for each single noun are combined. By selecting only the actual search terms, the processing time and accuracy imperfection of the recognition apparatus are compensated for, and the search terms are specified without stressing the user 210.

  In addition, the recognition result using the order search auxiliary database 282 and the recognition result using the reverse order search auxiliary database 283 are merged, whereby the accuracy of the recognition processing can be further improved, and the search time can be further increased. It can be shortened.

  FIG. 27 is a flowchart showing the operation of the fourth embodiment of the present invention.

  The fourth embodiment is an embodiment in which only reverse search is executed without executing the search.

  In FIG. 27, first, it is requested to input a search word by dividing it into single nouns (S230), and the number of words in which the single nouns constituting the compound noun are located in reverse order in the plural nouns is determined. “1” is set as the variable m to be shown, and “1” is set as the order q in the single noun sub-set Gbm-q constituting the reverse-order single noun set Gbm located at the reverse m-th word (S231). For each of “Yokosuka”, “Central”, and “Hotel”, recognition processing is performed for the reverse order noun sub-set H1-1, the reverse order noun sub-set H2-1, and the reverse order noun sub-set H3-1. Perform (S232, S233). Subsequently, for each single noun, a recognition result list is created in the case of performing reverse search by associating the recognition likelihood (S234). Then, influential constituent noun candidates are selected (S 235), an actual search word candidate list is created (S 236), and if the conditions for specifying the search words only by the confirmation process are satisfied (S 237), the search word candidates are given to the user. If it is presented (S238) and the response by the user is YES (S239), the search is terminated. On the other hand, if there is no candidate that satisfies the conditions for specifying the search word only by the confirmation process (S237), use Recognize in the range of the reverse most frequent noun subset.

  That is, in the fourth embodiment, for the single noun located in the reverse m-th word in the compound noun input by speech, the reverse no. M word (m is an integer value) of each compound noun registered in the database. This is a search device for compound nouns that are input by speech recognition in the range of a single noun set located in the reverse order m word that is a set of single nouns that are positioned.

  Specifically, the fourth embodiment is a search database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are registered by being divided into single nouns. And a reverse noun search auxiliary database in which a single noun set located in the reverse m word, which is a set of single nouns located in the reverse m word (m is an integer value) of each registered compound noun, is registered. When the user finishes inputting the compound noun, which is the search word, for each single noun, the single noun located in the reverse m word of the composite noun input is positioned in the reverse m word. Recognize in the range of a single noun set and create a reverse recognition result list in which pairs of constituent noun candidates that are the recognized single nouns and recognition likelihoods for the constituent noun candidates are arranged in order of recognition likelihood Reverse recognition result list And a reverse recognition result list for each of the single nouns input by voice, and among the constituent noun candidates described in the reverse recognition result list, the recognition likelihood exceeds a predetermined threshold. A candidate for a noun candidate to be selected as a candidate for a noun-possible component noun and a candidate for noun-possible noun component noun selection means and a combination of the above-mentioned noun-possible component noun candidates are subjected to a predetermined calculation on the recognition likelihood of the nominative noun candidate candidate. A search word recognition likelihood by calculating a search word recognition likelihood and creating a real search word candidate list arranged in descending order of the calculated search word recognition likelihood; and the real search word candidate Of these, a search term for executing a necessary dialogue with the user and confirming a search term for an existing search term having a search word recognition likelihood equal to or greater than a predetermined threshold. It is a search apparatus compound nouns which are voice input and a selection dialogue means.

  The above embodiment can be understood as an embodiment of the program. In other words, in the above embodiment, a compound noun composed of a plurality of single nouns is registered as a search word, and the search word is registered by being divided into each single noun, and the above registered A single noun set located in the reverse m-th word, which is a set of single nouns located in the reverse m-th word (where m is an integer value) of each compound noun, using a registered reverse order search auxiliary database, The procedure for searching for an input compound noun and when the user finishes inputting the compound noun as the search word for each single noun, the single noun located in the reverse m word of the input compound noun A pair of constituent noun candidates that are recognized as the recognized single noun and the recognition likelihoods for the constituent noun candidates are arranged in the order of recognition likelihood. Reverse order recognition A reverse recognition result list creation procedure for creating a list, and the reverse recognition result list are created for each of the single nouns input by speech, and among the constituent noun candidates described in the reverse recognition result list, predetermined The probable constituent noun candidates for each of the combinations of the reverse influential constituent noun candidates and the reverse influential constituent noun candidate selection procedure for selecting the constituent noun candidates having the recognition likelihood exceeding the threshold of Real search word candidate list creation that calculates a search word recognition likelihood by performing a predetermined calculation on the recognition likelihood of the search word, and creates a real search word candidate list arranged in descending order of the calculated search word recognition likelihood Among the actual search word candidates, the actual search word having a search word recognition likelihood equal to or higher than a predetermined threshold is necessary with the user. Run talk is an example of a program for executing a search word candidate selection dialogue procedure for determining the search term to the computer.

  Moreover, the said Example can be grasped | ascertained as an Example of the recording medium with which the said program is recorded. As the recording medium, FD, CD, DVD, HD, semiconductor memory, and the like are conceivable.

  It should be noted that the corporate name determination work performed in the above-described embodiment can be widely applied to the input interface portion in the product delivery service, telephone number search, postal code search, and the like.

  According to the above embodiment, not only the order of composition from the top is taken into account, but the frequent noun used at the end, the frequent noun used at the second word from the end, By counting the words, even if they are used at the same end, if they are counted from the beginning depending on the number of constituent words, the words that are not recognized as the most frequently used words can be caught as priority recognition objects. .

  In the above embodiment, the search word recognition likelihood is the sum of the recognition likelihoods of each single noun as shown in FIG. 26. However, when calculating the search word recognition likelihood, You may make it multiply recognition likelihood.

  If the search word recognition likelihood of the search word candidate in the actual search word candidate list is equal to or greater than the threshold (for example, the threshold is 280), the search word is determined only by the confirmation process by the user 210. Conversely, if the search word recognition likelihood is less than the threshold, the search word cannot be determined only by the confirmation process by the user 210.

  In preparation for the case where the search term cannot be determined as described above, processing such as recognition of a single noun and calculation of recognition likelihood in the range of the single noun sub-set Gb2-p and the reverse single noun sub-set H2-q are performed. Alternatively, it may be executed in advance. In other words, in parallel with the process of presenting the search word candidate to the user 210 on the display and requesting confirmation (S224) and the response by the user 210 (S225), the single noun sub-set, the reverse order single Single noun recognition, recognition likelihood calculation (S202, S212), re-selection of influential constituent noun candidates (S204, S214), and actual search word candidate list update (S222) are executed within the noun sub-set range.

  In this way, since the user 210 cannot see the processing status inside the search device 201 at all, the user 210 can search for the search terms registered in the search database 281 all at once. It seems as if the recognition process is being performed at high speed.

  FIG. 28 is a block diagram showing a spoken dialogue compound noun search apparatus 301 according to the fifth embodiment of the present invention.

  This speech interactive compound noun search device 301 includes a speech input unit 302, a speech recognition unit 303, speech recognition software 303S, a speech recognition result output unit 304, an actual search word candidate list creation unit 305, a dialogue The controller 306, the audio output unit 307, and the audio output software 307S are configured.

  In the compound noun search device 301 inputted by voice, the voice of the user P inputted through the voice input unit 302 is sent to the voice recognition unit 303, and the voice recognition unit 303 performs voice recognition processing on the input voice. In addition, the system database 308 is used. Further, the voice recognition unit 303 performs a recognition process on the voice input by the user P using the voice recognition software 303S.

  The system database 308 includes a search database 381, a search auxiliary database 382, and a YES / NO database 383.

  The search database 381 is a database in which compound nouns composed of a plurality of single nouns are registered as search words, and the search words are registered by being divided into single nouns.

  In the search auxiliary database 382, a group of single nouns written in the n-th (n is an integer value) of each registered compound noun is referred to as an n-th single noun set group. A group of nouns is registered in order of frequency, and single nouns registered in order of frequency are collected in order from the highest frequency for every predetermined number (500 in the search device 301). And a database in which a plurality of sub-set groups are formed.

  The YES / NO database 383 is a database that recognizes the contents (for example, yes / no, yes / no) that the user P has responded to.

  The speech recognition software 303S selects the search database 381 or the search auxiliary database 382 from the system database 308 in accordance with the processing scene of the search device 301.

  When the search term is inputted by voice, the search auxiliary database 382 is referred to, and when the response to the correctness confirmation to the user P is recognized, the YES / NO database 383 is referred to.

  The voice recognition unit 303 uses the voice recognition software 303S in the voice recognition process, and the voice output unit 307 uses the voice output software 307S in the voice output.

  When the user inputs a compound word, which is a search word, for each single noun, the speech recognition result output unit 304 recognizes the single noun of the nth notation of the compound noun in the single noun set group of the nth notation. In addition, a recognition result list is created by performing recognition processing within the range of the first sub-set group including the most frequent single noun among the n-th notation single-noun set group, and corresponding the recognition likelihood. It is an example of a recognition result list creation means.

  Further, the speech recognition result output unit 304 displays a recognition result list in which pairs of constituent noun candidates that are the recognized nouns and recognition likelihoods for the constituent noun candidates are arranged in order of recognition likelihood. A constituent noun that has a recognition likelihood that exceeds a predetermined first threshold among the constituent noun candidates that are created for each single noun that is input by speech and that is described in the recognition result list. It is an example of the powerful constituent noun candidate selection means selected as a noun candidate.

  The real search word candidate list creation unit 305 includes a real search word list creation unit 351 and a partial match search word list creation unit 352.

  The real search word list creation unit 351 is a part that extracts a real search word from a search word stored in the search database 381 based on the voice recognition result output by the voice recognition result output unit 304.

  When the likelihood of only some single nouns out of the plurality of single nouns constituting the search word exceeds the first predetermined threshold, the partial match search word list creation unit 352 sets the first threshold. Partial match search word candidate extraction means for extracting from the search database partial match search word candidates comprising the single nouns having a likelihood exceeding that and having the same number of single nouns as the search words It is an example.

  The dialogue control unit 306 includes a search word candidate selection dialogue unit 361 and a partial match search word candidate selection dialogue unit 362.

  The partial match search word candidate selection dialogue unit 362 calculates the partial match likelihood by calculating the likelihood of each single noun constituting the extracted partial match search word candidate by a predetermined calculation method. It is an example of a means. Note that the predetermined calculation method for calculating the partial match likelihood is a method of adding the likelihood of each single noun in the search device 301, but it may be multiplied instead of the addition.

  The audio output unit 307 outputs audio using audio output software 307S.

  Note that the speech recognition unit 303 performs speech recognition processing for the partial match search word candidate for which the calculated partial match likelihood exceeds the predetermined second threshold. It is an example of a means.

  Next, the search device 301 will be described more specifically.

  In the following description, a case will be described in which a speech noun search device 301 is provided in an input interface portion of a service for determining a corporate name such as a company name.

  There are as many as 22 million corporate names listed in phone books across Japan, and it takes a considerable amount of time to identify one corporate name out of 22 million based on the input voice. In the current speech recognition technology, the corporate name can be recognized and processed within the real time of dialogue processing (when the search device and the user P interact, the time when the user P does not feel stress or unnaturalness). In this case, the recognition accuracy is very low because there are many very similar data.

  FIG. 29 is a diagram showing a specific example in which the corporate name in the search database 381 is composed of compound nouns in the search device 301, and these compound nouns are registered separately for each single noun constituting the compound nouns.

  The frequency of each noun constituting the corporate name registered in the search database 381 is examined, and single nouns are recorded in order of frequency, and the search auxiliary database 382 is created.

  The total number of single nouns constituting the 22 million corporate names is 6.5 million. In addition, the total number of single nouns (single nouns located in the first word in the corporate name) of nouns constituting 22 million corporate names is about 3.6 million. The total number of single nouns (single nouns located in the second word in the corporate name) is about 2.5 million, and the third single noun (single noun located in the third word in the corporate name) ) Is about 2.7 million types, and the total number of single nouns in the fourth notation (single noun located in the fourth word in the corporate name) is about 1 million types, ... (the maximum number of constituent words is 7) In the real time of dialogue processing, it is impossible to recognize each single noun with respect to one compound noun, and it is expected that the accuracy is low.

  FIG. 30 is a diagram illustrating a specific example of data stored in the search auxiliary database 382 in the search device 301.

  The single nouns in the search auxiliary database 382 are single nouns constituting the compound nouns registered in the search database 381, and a single noun set group is formed for each notation order in the compound nouns. By collecting only the single nouns whose notation order is the first in the compound noun, the first noun single noun set group Gc1 is formed, and in the first noun single noun set group Gc1, the single nouns are arranged in descending order of frequency of use. In addition, these registered single nouns are grouped into 500 single nouns in order from the highest frequency to form a plurality of sub-set groups, and the first sub-set in descending order of frequency. A group Gc1-1, a second sub-set group Gc1-2,..., An m-th sub-set group Gc1-m are formed.

  The first subset group Gc1-1 including the single noun having the highest usage frequency is a target to be recognized with priority over the other subset groups Gc1-2,..., Gc1-m.

  In addition, only single nouns with the second notation order in the compound noun are collected to form a second noun single noun set group Gc2, and in the second noun single noun set group Gc2, in descending order of frequency of use. Single nouns are registered, and these registered single nouns are grouped in units of 500 single nouns in descending order of their frequency, and a plurality of sub-set groups are formed. A first sub-set group Gc2-1, a second sub-set group Gc2-2,..., An m-th sub-set group Gc2-m are formed.

  The first subset group Gc2-1 including the single noun having the highest use frequency is a target to be recognized with priority over the other subset groups Gc2-2,..., Gc2-m.

  Hereinafter, in the same manner as described above, only the single nouns with the nth order of compound nouns are collected to form the nth single noun set group Gcn. In the nth single noun set group Gcn, Single nouns are registered in descending order of frequency of use, and the registered single nouns are grouped in units of 500 single nouns in order from the highest frequency to form a plurality of sub-set groups. The first sub-set group Gcn1, the second sub-set group Gcn2,..., The n-th sub-set group Gcn-m are formed in descending order.

  The first subset group Gcn-1 including the single noun having the highest usage frequency is a target to be recognized with priority over the other subset groups Gcn-2,..., Gcn-m.

  By the way, if the number of nouns constituting one sub-set group is N, N = 500 in the search device 301, and this number is determined as follows. That is, a single noun that can be processed by the speech recognition software 303S within the real time of dialogue processing (when the search device and the user P interact, it is a time during which the user P does not feel stress or unnaturalness). When the number is T and the average number of single nouns of all search words registered in the search database 381 is M, N = T / M.

  Specifically, it is assumed that the number T of single nouns that can be processed by the speech recognition software 303S within the interactive processing real time is 1500, and the average number of single nouns M of all search words registered in the search database 381 is 3. N = T / M = 1500/3 = 500.

  Note that the number N of single nouns constituting each of the sub-groups Gc1-1,..., Gcn-m may be a number other than 500. It is determined by the number T of single nouns (the performance of the speech recognition software 303S) and the average number M of single nouns included in the search terms registered in the search database 381.

  Next, the operation of the search device 301 will be described.

  FIG. 31 is a flowchart specifically illustrating a processing process for determining a search term in the search device 301.

  As this premise, it is assumed that a search database 381 and a search auxiliary database 382 have been created.

  First, the user P is requested to input a corporate name (compound noun) to be searched in a form divided into single nouns (S300). Then, the function k of the sub-group in each single-noun group related to the compound noun that is the search word (function indicating the rank of the sub-group) is set to 1. That is, if the function k = 1 of the subset group, each single noun is recognized using the subset groups Gc1-1, Gc2-1, Gc3-1,. Is done.

  Consider the case where the compound noun “Yokohama / Tokyu / Hotel” is input as a search term by voice.

  First, for the single nouns “Yokohama”, “Tokyu”, and “Hotel” input by speech, the first sub-noun group Gc1-1 and the second-notation single noun set in the first-nominated single noun set group Gc1, respectively. The speech recognition unit 303 performs a recognition process using the first sub-set group Gc2-1 in the group Gc2 and the first sub-set group Gc3-1 in the third-named single noun set group Gc3 (S301). That is, the single noun “Yokohama” input by speech is recognized within the range of the first sub-group Gc1-1, and the range of the first sub-set group Gc2-1 for the single noun “Tokyu” input by speech. The single noun “hotel” input by voice recognition is recognized and processed within the range of the first sub-group Gc3-1. Then, the speech recognition result output unit 304 outputs the recognition result.

  Of the single nouns “Yokohama”, “Tokyu” and “Hotel” input by voice, the single noun “Yokohama” is included in the first sub-group Gc1-1 as shown in FIG. It is assumed that the noun “hotel” is included in the first sub-group Gc3-1. However, since the frequency rank of the single noun “Tokyu” is the frequency rank 951 as shown in FIG. 30, it is not included in the first subset group Gc2-1 and included in the second subset group Gc2-2. It is.

  FIG. 32 is a diagram showing a specific example of a recognition result list in which the recognition result for each single noun constituting the search word (compound noun) input by speech and the recognition likelihood correspond in the search device 301. is there.

  When one single noun is input by speech, the speech recognition result output unit 304 outputs a plurality of constituent noun candidates for the input single noun and the recognition likelihood for each of the plurality of constituent noun candidates. The recognition result list is created by associating the outputted constituent noun candidates with the recognition likelihood (S301). A specific example of the recognition result list created in this way is shown in FIG.

  For example, when the single noun “Yokohama” is inputted by voice, the voice recognition unit 303 uses the inputted voice as “Yokohama”, “Yokosuka”, “Yokomizo”, “Yokoyama”,... The recognition likelihood of “Yokohama” is 95, the recognition likelihood of “Yokosuka” is 90, the recognition likelihood of “Yokomizo” is 81, and the recognition likelihood of “Yokoyama” is 75. is there.

  Note that the recognition likelihood for each single noun is individually determined by the speech recognition software 303S when the single noun is input by voice.

  Here, a constituent noun candidate having a recognition likelihood equal to or greater than a predetermined threshold is set as a dominant constituent noun candidate. In the search device 301, the predetermined threshold is set to 80. That is, a constituent noun candidate having a recognition likelihood of 80 or more is a dominant constituent noun candidate. That the predetermined threshold is 80 is set in advance when the databases 81 and 82 are created.

  Then, from each recognition result shown in FIG. 32, a single noun having a recognition likelihood of 80 or more is selected as a potential constituent noun candidate (S302).

  At this point, when there is a first sub-set group in which the recognition process has not yet been completed in each search auxiliary database 382, the speech recognition unit 2 restarts the recognition process for the first sub-set group, The recognition result output unit 3 repeats the process of outputting the recognition result and sending it to the search word candidate list creation unit 305.

  In the search word candidate list creation unit 305, the actual search word candidate list creation unit 3051 selects a potential constituent noun candidate based on the recognition result, creates all possible combinations of potential constituent noun candidates, and refers to the search database 381. On the other hand, a real search word candidate list obtained by extracting real search words is created (S303) and sent to the dialogue control unit 306.

  The actual search word candidate list is a list of search word candidates that are presented to the user P and asked for confirmation among the search words existing in the search database 381.

  By the way, the frequency rank of the second noun single noun “Tokyu” is ranked 951 in the second noun single noun set group Gc2 stored in the auxiliary database 82 as shown in FIG. The first noun set group Gc2-1 of the single noun set group Gc2 does not include the single noun “Tokyu”. Therefore, the single noun “Tokyu” is not included in the recognition result for the voice-input single noun “Tokyu” shown in FIG.

  Therefore, the actual search word candidate list is not created in the step of creating the actual search word candidate list (S303).

  At the same time, the partial match search word list creation unit 352 calculates partial match likelihoods in the partial match search word candidates (S331), extracts partial match search word candidates from the search database 381, and sends them to the dialogue control unit 306. send.

  The partial match search word candidate is a likelihood that exceeds the first threshold when the likelihood of only some of the single nouns constituting the search word exceeds a first predetermined threshold. This is a search word candidate that includes the single nouns having the same number of single nouns as the search words.

  The partial match likelihood is a likelihood obtained by calculating the likelihood of each single noun constituting the extracted partial match search word candidate by a predetermined calculation method.

  The dialogue control unit 306 executes a search word candidate selection dialogue using the actual search word candidates and the partial match search word candidates.

  That is, if the search word candidate in the actual search word candidate list satisfies the specified conditions of the search device that can specify the search word only by the confirmation process with the user, the voice output unit 307 outputs the confirmation process guidance output command. Send to.

  On the contrary, if the search word candidates in the real search word candidate list do not satisfy the conditions for specifying the search word only by the confirmation process with the user (S321), or if the real search word candidates are not extracted, Using the partial match search word candidate, the speech recognition unit 303 performs speech recognition processing again on the speech initially input to the speech input unit 1 (S332), and the recognition likelihood is calculated.

  When the calculated recognition likelihood satisfies the specified state of the search device that can specify the search word only by the confirmation process with the user (S321), an output command of the confirmation process guidance is sent to the voice output unit 307. The search word candidate is presented, and the user P is asked for confirmation (S322).

  If the search word candidate that can be specified only by the confirmation process is not extracted from the actual search word candidate or the partial match search word process using the partial match search word candidate (S321), the recognition process is terminated at this point. Then, for the constituent noun set group having the second highest usage frequency (S311), the effective constituent noun candidates obtained as a result of the recognition process are used (S312 and S313), and the real search word candidate list and the partial match search word list are obtained. Update (S314), and repeat the search word candidate selection dialogue and the partial match search word processing (S315).

  When the voice output unit 307 receives the confirmation processing guidance output command, the voice output unit 307 outputs a guidance for confirming whether or not the candidate is designated as a search word candidate by the user.

  Then, when a response to the confirmation processing guidance is input again from the voice input unit 302, the voice recognition unit 303 refers to the YES / NO database 383, recognizes the user's response, and from the voice recognition result output unit 304. When the recognition result is output and a response indicating affirmation can be obtained from the user (S315), the dialogue control unit 306 outputs a command for guidance to the user that the search term specification has been completed. Send to 307.

  Until the search term can be identified, the search word candidate list is updated (S314), the search word candidate selection dialogue, and the partial match search word processing (partial match search word candidates are extracted from the search database 381, and this partial match is extracted. The process of calculating the partial match likelihood that is the total likelihood of the single nouns constituting the search word candidate is repeated.

  When the recognition process for all the constituent noun sets in the search auxiliary database 382 is completed and the actual search word candidate list cannot be updated any more (S315), the dialog control unit 306 is higher in the real search word candidate list. Then, a guidance output command for presenting search word candidates to the user is sent to the voice output unit 307.

  That is, the search device 301 updates at least one set of constituent noun sets other than the priority recognition target noun group (set 1 consisting of a specified number) during the update of the actual search word candidate list and the search word candidate selection dialogue. The number of constituent nouns is defined so that the recognition process is completed.

  Next, the operation of the search device 301 will be described more specifically.

  Using these influential constituent noun candidates, the actual search word candidate list creation process is performed. For the first sub-group Gc2-1 relating to the second notation single noun, the influential constituent noun candidate relating to “Tokyu” is selected. Therefore, the search device 301 performs a partial match search process as soon as the recognition process for the first sub-groups Gc1-1 and Gc3-1 is completed.

  That is, the search which has the prominent constituent noun candidate "Yokohama" or "Yokosuka" in the 1st subgroup Gc1-1 in the 1st single noun group Gc1 which is the 1st single noun in the 1st single noun A candidate word “hotel” or “hall” in the first sub-set group Gc3-1 in the third-notation single-noun set group Gc3 that is the third single noun is designated as the third single noun. Search word candidates for nouns are searched from the search database 381.

  FIG. 33A is a diagram illustrating an example of partial match search word candidates and partial match likelihoods in the search device 301.

  In the search device 301, “Yokohama / Grand / Hotel”, “Yokohama / Tokyu / Hotel”, “Yokohama / Tobu / Hotel”, “Yokosuka / Central / Hotel”, “Yokohama / Central / Hall”, “Yokohama / Grand / Hall” ”,“ Yokohama / Music / Hall ”,“ Yokohama / Chuo / Hall ”, and“ Yokosuka / Art / Hall ”are searched as partial match search word candidates.

  In the above example, for example, among “Yokohama / Grand / Hotel”, “Yokohama” and “Hotel” are potential constituent noun candidates as shown in FIG. 32, and “Grand” is not a potential constituent noun candidate. . In other words, in “Yokohama / Grand / Hotel”, only some simple nouns are possible constituent noun candidates, and therefore “Yokohama / Grand / Hotel” is a partial match search word candidate.

  For example, in “Yokohama / Grand / Hotel”, the recognition likelihood of “Yokohama” is 95 as shown in FIG. 32, and the recognition likelihood of “Hotel” is 88 as shown in FIG. Yes, 183 which is a value (95 + 88 = 183) obtained by adding the respective recognition likelihoods of these prominent constituent noun candidates is the partial match likelihood.

  In the search device 301, it is assumed that the threshold value (second threshold value) in the partial match likelihood is 130.

  Then, the speech recognition process for the first input search word is executed again by the speech recognition unit 303 with these nine candidates as recognition targets.

  As a result, as shown in FIG. 33 (2), the search word recognition likelihoods of “Yokohama / Tokyu / Hotel” and “Yokohama / Tobu / Hotel” are 95 and 81, respectively. The threshold value) exceeds 80. Therefore, “Yokohama / Tokyu / Hotel” and “Yokohama / Tobu / Hotel” are extracted as search word candidates. The search term identification is completed by confirming the presentation from the top.

  In the search device 301, the criteria for determining whether or not the search word can be determined only by the confirmation process with the user for the result of the partial match search word processing are the threshold value specified in advance by the compound noun search device and the search word recognition likelihood. When the threshold value (third threshold value) for the re-recognized result is 80, the search word candidate having the search word recognition likelihood exceeding the third threshold value 80 is only confirmed. It is determined that the condition that can determine the search term is satisfied.

  Therefore, in the search device 301, after re-recognition, the search word candidates that satisfy the search word determinable condition only by the confirmation process are “Yokohama / Tokyu / Hotel”, “Yokohama / Tobu / Hotel.

  In the search device 301, not only the remaining constituent nouns are recognized using the real search word list creation process and the search word candidate selection dialogue, but also the partial match search words using the partial match likelihood. Process.

  As a result of the partial match search word processing, a search word candidate selection dialogue is performed using the calculated partial match likelihood. When there is no search word candidate having a partial match likelihood exceeding the specified threshold (third threshold), or when there is no search word candidate to be selected as a result of the partial match search word processing, the second sub-set Continue the actual search word list creation process and search word candidate selection dialogue using the results recognized for the group. At the same time, while updating the powerful constituent noun list, the partial match search word processing and the search word candidate selection dialogue for the result are repeated.

  In the search device 301, since the single noun “Tokyu” is included in the second subset group Gc2-2, it is selected when the real search word list creation process using the result recognized for the second subset group is performed. Search term candidates exist. On the other hand, as a result of the partial match search word processing, a search word candidate selection dialogue is performed using the calculated partial match likelihood, and there is a search word candidate having a partial match likelihood exceeding a specified threshold (third threshold). To come. In this case, these two results may be merged.

  When the search device 301 performs the real search word list creation process based on the recognition result for the set group of the constituent nouns for each constituent order, at the same time, partial information on the potential constituent noun candidates that satisfy the conditions predetermined by the search device Create a partial match search term list using That is, the partial match likelihood is calculated by using the prominent constituent noun candidates, the search word having the partial match likelihood equal to or higher than the predetermined threshold is extracted, the extracted search word is determined as the recognition target, and is input first. The recognition process of the search term is executed again to calculate the recognition likelihood.

  When it is determined that the calculated recognition likelihood satisfies the condition for specifying the search word only by the confirmation process with the user, a search word candidate selection dialogue is performed. When there is no search word candidate having a partial match likelihood equal to or greater than the specified threshold value, or when no search word candidate to be selected exists as a result of the search word candidate selection dialogue, for each remaining nth component noun set group As a result of the recognition process, the updated effective constituent noun candidates are used, and the actual search word list creation process, the search word candidate selection dialog, the partial match search word processing, and the search word candidate selection dialog are repeated.

  As described above, by using the partial information of the constituent nouns, the search process can be executed accurately and quickly without being influenced by the bias of the constituent nouns.

  By the way, in practice, all the single nouns constituting the search word are rarely frequent, and all the single nouns constituting the search word are rarely infrequent, and the frequency is biased. There are many cases, and among each single noun, single nouns that are used frequently are set in the first sub-set group, so they are output as the first recognition processing result, but single nouns that are used less frequently Is not output as a recognition result until the recognition processing for the subset group including the word is performed. Therefore, a combination of each constituent noun candidate having a likelihood equal to or greater than a predetermined threshold is created, and an actual search word candidate is selected. Even if it is extracted, there is no correct answer.

  Therefore, the search device 301 partially matches a search word partially including a prominent constituent noun from large vocabulary search words that cannot be recognized and processed correctly in real time by setting all of them as recognition targets. By extracting from the search database 381 as search word candidates, it is possible to narrow down potential candidates. By this narrowing down, search processing can be performed without causing the user to wait.

  That is, according to the search device 301, by using partial information of single nouns together, it is possible to narrow down the correct search word and further improve the recognition accuracy and the recognition speed.

  In the flowchart shown in FIG. 31, after step S303, apart from the route proceeding to step S321, the route proceeding to step S311 and the route proceeding to step S331 are indicated by broken lines. This means that the routine of steps S321 to S323, the routine of steps S311 to S315, and the routes of steps S331 and S332 are processed in parallel.

  Further, the search device 301 can be grasped as a program invention.

  That is, the search device 301 has a likelihood that exceeds the first threshold when the likelihood of only some of the single nouns constituting the search word exceeds the first predetermined threshold. A partial match search word candidate extraction procedure for extracting a partial match search word candidate from the search database, the partial match search word candidate having the same number of single nouns as the single noun constituting the search word, and the extracted The partial match likelihood calculation procedure for calculating the partial match likelihood by calculating the likelihood of each single noun constituting the partial match search word candidate by a predetermined calculation method, and the calculated partial match likelihood It is an example of the program which makes a computer perform the speech recognition processing procedure for partial matching search word candidates which performs a speech recognition process about the search word candidate exceeding the 2nd threshold value.

  Further, the search device 301 includes a search database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are registered by dividing each single noun into the search nouns. A group of single nouns written in the nth (n is an integer value) of each compound noun is called an nth single noun set group, and the nth single noun set group has a high frequency. A single noun that is registered in order and that is registered in order from the highest frequency is collected into a predetermined number in order from the highest frequency, and a search auxiliary database in which a plurality of sub-set groups are formed A compound noun search program, when a user inputs a compound word that is the search word for each single noun, the n-th notation of the compound noun Recognized by a single noun group In addition, a recognition result list is created by performing recognition processing within the range of the first sub-set group including the most frequent single noun among the n-th notation single-noun set group, and corresponding the recognition likelihood. The speech input is a recognition result list in which a set of a recognition result list, a constituent noun candidate that is a single noun subjected to the recognition processing, and a recognition likelihood for the constituent noun candidate are arranged in order of recognition likelihood. Among the constituent noun candidates that are created for each single noun and are listed in the recognition result list, a constituent noun candidate having a recognition likelihood exceeding a predetermined first threshold is set as a potential constituent noun candidate. When the probable constituent noun candidate selection procedure to be selected and the likelihood of only some of the single nouns constituting the search word exceeds the first predetermined threshold, the first threshold is set to Has a likelihood that exceeds A partial match search word candidate extraction procedure for extracting partial match search word candidates from the search database, the partial match search word candidates having the same number of single nouns as the search words. The partial match likelihood calculation procedure for calculating the partial match likelihood by calculating the likelihood of each single noun constituting the partial match search word candidate by a predetermined calculation method, and the calculated partial match likelihood This is an example of a program that causes a computer to execute a partial matching search word candidate speech recognition processing procedure for performing speech recognition processing on the partial matching search word candidate exceeding the second threshold.

  It should be noted that the corporate name determination work performed by the search device 301 can be widely applied to the input interface portion in the product delivery service, telephone number search, postal code search, and the like.

  FIG. 34 is a block diagram showing a voice interactive voice interactive search device 401 according to the sixth embodiment of the present invention.

  The voice interactive search device 401 includes a voice input unit 402, a voice identification unit 403, a dialog control unit 404, a voice output unit 405, voice output software 405S, a voice recognition device 406, a system database 408, It is constituted by.

  The voice identification unit 403 includes a voice recognition unit 431 and a recognition result adjustment unit 432.

  The voice recognition unit 403 uses a voice recognition device 406, and the voice output unit 405 uses voice output software 405S. Further, the speech recognition processing for the input speech by the speech identification unit 403 and the dialogue control unit 404 use the system database 408.

  The system database 408 includes a search information database 481 in which search information to be searched is recorded, a beat database 482, and a YES / NO database 483.

  FIG. 35 is a diagram showing an example of the entire image of the search information database 481 used in the voice interactive search device 401. As shown in FIG.

  The search information database 481 stores search information itself as a search target.

  FIG. 36 is a diagram showing an example of the beat database 482 used in the voice interactive search device 401.

  In the beat database 482, search information to be searched is divided into a first beat, a second beat, a third beat,... By a predetermined method. Then, only the first beat is extracted, the extracted first beats are arranged in order of use frequency, the second beat following each first beat is arranged in order of frequency of use, and the third beat following each second beat is used. The beats are arranged in the order of frequency, and the beats arranged are stored in the beat database 482.

  When the search information database 481 is configured by a large vocabulary that cannot be processed within the dialog processing real time, a predetermined number is set as the number that can be processed within the predetermined dialog processing real time for each beat constituting the beat database 482. Determine the number of

  The voice identification unit 403 uses the voice recognition device 406 to perform recognition processing for user input. The speech recognition device 406 selects a database to be recognized from the system database 408 in accordance with the processing scene of the search device. That is, when the user inputs search information, the beat database 482 is referred to, and when the user responds to the candidate presentation, the YES / NO database 483 is referred to.

  When the user inputs search information to be searched, the input search information is sent to the voice identification unit 403. In the voice identification unit 403, the voice recognition unit 431 recognizes the first beat stored in the beat database 482 and sends the recognition processing result to the recognition result adjustment unit 432. Based on the recognition result of the first beat, the recognition result adjustment unit 432 extracts a leading first beat that is the first beat having a score equal to or higher than a predetermined score, and sends information to the voice identification unit 403.

  The voice recognition unit 431 recognizes the second beat candidate having a high frequency linked to each of the leading first beats as a second beat recognition target, and similarly recognizes the third beat, the fourth beat, .., And narrow down candidates while exchanging information with the recognition result adjustment unit 432 until the last beat.

  The dialogue control unit 404 calculates the total score by adding the recognition result score of each beat to the received recognition result, and the calculated total score can be confirmed only by the confirmation process by the user. If the search device 401 determines that the score is reached, the voice output unit 2 is instructed to confirm the presentation to the user.

  The search is successful when the user is prompted to input a correct / incorrect response to the presentation and the response of Yes is recognized.

  When it is determined that it is not worthy of the confirmation processing, based on the recognition result of each beat stored in the voice identification unit 403, a voice output unit is provided to present the first beat in the order of recognition score. In step 405, presentation of the first beat candidate is instructed.

  When a correct / incorrect response to the presentation of the first beat is requested and a response of “Yes” is recognized, the voice identification unit 403 recognizes the second beat following the confirmed first beat in a descending order of frequency of use. In the same manner as described above, the search information as the search target is determined while narrowing down the recognition targets sequentially, such as determining the second beat.

  When the user responds to the presentation confirmation, the voice recognition unit 431 recognizes the user response with the YES / NO database 483 as a recognition target.

  Next, the information search operation when the voice interactive search device 401 is applied to the input interface whose task is to determine the personal surname will be described in detail.

  The search information in this case is a personal surname. There are 180,000 personal surnames in Japan.

  FIG. 37 is a diagram showing a specific example of the search information database 481 in the voice interactive search device 401.

  In this specific example, when an individual surname is divided by a beat, when the surname is expressed in Kanji, it is divided so that the reading of one Kanji character is counted as one beat.

  FIG. 38 is a diagram showing an example of the beat database 482 in the voice interactive search device 401.

  There are 3000 types of first beats, which are stored in the beat database 482 in descending order of frequency of use. As shown in FIG. 39, the second beat following each of the first beats is ordered in descending order of frequency of use. Similarly, for each second beat, the third beat following the second beat is stored in order of frequency of use, and so on, and the beat database 482 is created.

  By the way, the current state of speech recognition technology requires enormous recognition processing time when batch recognition processing is performed for 180,000 individual surnames, and there are many very similar surnames, so the recognition accuracy is very low. In particular, for 180,000 recognition targets, it is known that it is impossible to end the recognition process and return a response with a certain degree of accuracy within a time when the user does not feel unnatural. .

  In the voice interactive search device 401, when extracting a promising candidate from the result of each recognized beat with respect to search information inputted by speech, the recognition score is compared with a threshold value, and a candidate having a recognition score exceeding the threshold value is selected as a promising candidate. Determined as a candidate. Similarly, when determining whether or not the search information can be specified only by the user confirming process, if the total score calculated for the search information to be searched exceeds the threshold, the user It is determined that the search information can be specified only by performing the confirmation process.

  Next, the voice interactive search device 401 will be described more specifically.

  FIG. 39 is a flowchart showing an operation of determining search information in the voice interactive search device 401.

  An operation when the user inputs “Matsumoto” by voice (S401) and the inputted “Matsumoto” is confirmed will be described.

  For voice input by the user, the leading beat candidate in the beat database 482 is preferentially recognized and processed in descending order of frequency of use (S402).

  In FIG. 40, in the voice interactive search device 401, the first beat candidates in the beat database 482 corresponding to the first beat constituting the search information voice-input by the user are arranged in order of frequency of use (the score is highest). It is a figure which shows the recognition result after carrying out recognition processing preferentially among a predetermined number of first beat candidates among the arranged first beat candidates.

  Here, when it is determined that the threshold value as a leading first beat is 80, “mas”, “matsu”, and “masa” are extracted as leading leading beats (S403).

  The second beat candidate connected to each leading first beat is set as a recognition target (S404), the second beat candidates are arranged in the order of frequency of use (the score is large), and among the arranged second beat candidates, A predetermined number of second beats are preferentially recognized (S405).

  The extracted second beat candidates in the voice interactive search device 401 are shown in FIG.

  Connected to the first beat candidate (first beat) “mas”, the second most frequent beat candidate is “da”, the highest priority, and the second leading beat candidate “matsu” The second beat candidate with the highest frequency is “Moto”, which leads to “Masa”, the third first beat candidate, and the second beat candidate with the highest frequency is “Moto”. The second beat candidate with the second highest frequency is “Ta”, and the second beat candidate with the second highest frequency is “Zaka”. Then, recognition processing for the second beat candidate is performed.

  If the end of the search information input by voice is not recognized, the recognition process is subsequently executed (S406).

  FIG. 41 is a diagram illustrating an integration result (total score) of a score that is a recognition result of the leading beat candidate and a score that is a recognition result of the second beat candidate in the voice interactive search device 401.

  These results are sent to the dialogue control unit 404, and the total score of the score of the first beat and the score of the second beat is calculated (S407), and it is determined whether it can be determined only by the confirmation process (S411). . In the above example, the recognition score of each beat is added to obtain the total score, and the threshold value for determining that the search information can be confirmed only by the confirmation process by the user is set to 160, and the total score exceeds the threshold value 160. Search information candidates are extracted.

  As a result, among the search information candidates “Matsumoto”, “Masamoto”, “Matsuto”,..., The search information candidate exceeding the threshold 160 is one “Matsumoto” candidate, and this search information candidate “ “Matsumoto” is confirmed and presented to the user (S412). When a Yes response is obtained from the user (S413), the search and determination are completed.

  According to the speech interactive search device 401, since the frequency of use of beats constituting the search information is taken into consideration and the ease of connection between the beats is taken into consideration, even if the recognition target is a large vocabulary, It can be processed in time. In addition, by performing recognition processing in parallel for each number of recognition targets that can be answered with an accuracy that is considered acceptable for the user (S421 to S427, S431), user stress due to waiting time, user due to misrecognition The stress is reduced.

  There are many types of personal surnames, and there are many similar candidates. However, the recognition of personal surnames can be used in a wide range of customer care such as call center accepting operations and product delivery services.

  The voice interactive search device 401 decomposes the search information, which is the recognition target vocabulary, for each beat and performs recognition processing for each beat. At this time, not all beats are handled uniformly, Based on the prediction that there is a high possibility of utterance, the more frequently used ones are recognized with priority. In this case, the priority method may be weighted on the recognition result. .

  In addition, for the second beat and the third beat using the connectability between beats and the connection frequency, only the next connected beat is arranged in order of the highest use frequency for each previous beat, By performing recognition processing in order from the previous beat, only the second beat recorded for the first beat extracted as the leading first beat is prioritized in descending order of frequency of use based on the recognition result for the previous beat. A rank is assigned, it is determined as a recognition target, and recognition processing is performed. Similarly, the recognition processing is continued until the last beat such as the third beat, the fourth beat,.

  In the unlikely event that the leading candidate is mistakenly extracted and the correct candidate cannot be derived, by confirming in order from the leading beat, it has the function of advancing the narrowing dialogue using the hierarchical structure of the beat, and the search information itself is recognized and processed Can be expected to narrow down candidates more efficiently and efficiently.

  If the search information database consists of large vocabularies that cannot be processed within the real time of interactive processing, the search information database can be processed in real time, and a response with an accuracy acceptable to the user can be returned. The number of objects that can be defined in advance is determined as a priority recognition object in descending order of frequency of use, and recognition processing is performed.

  Once a potential leading beat candidate has been extracted, the specified number that can be processed in real time is determined as a priority recognition target for the second beat connected to it, and recognition processing is performed similarly for the third and fourth beats. Advance the process, calculate the total score, and when the recognition process for the first recognition target first beat is completed, start the recognition process for the specified number that can be processed in real time from the first frequently used first beat, If a new leading first beat candidate is extracted from there, the processing is continued for the second and third beats connected to it.

  At the time when the total score can be calculated, if there is a candidate that can be determined to be determined only by confirmation processing by the user, the candidate that can be determined by performing presentation confirmation and only confirmation processing by the user Is not present, at that time, the recognition processing result for the recognition target list of the next most frequently used beat should have been output by parallel processing (S421 to S425). In S426), it is determined whether or not the confirmation process is to be performed (S411), and the process is repeated until it can be confirmed (S431), whereby the accuracy of the recognition engine for confirming the target information is compensated by the search method by the speech interactive information retrieval apparatus. is there.

  The user does not see the processing state inside the system at all, and the speech recognition process is uniformly performed on the search information in the search information database, and the result is presented. Therefore, without causing the user to feel unnaturalness and stress, the input search information can be recognized within an appropriate time, and the recognition result can be presented with high accuracy.

  There are the following methods for extracting beats. That is, in the case where the search information is written in hiragana, how to cut out one hiragana character as one beat (the roaring sound, the sound repellent, and the long sound are counted as one beat together with the previous hiragana), the search information In hiragana notation, starting from the beginning, for each predetermined number of hiragana characters, cut out to be 1 beat (stuttering, repelling, long sound is counted as one character together with the previous hiragana), depending on the number of constituent hiragana characters How to cut out the number of characters that make up one beat (for example, search information consisting of 4 characters counts as 1 beat every 2 characters from the top, search information consisting of 3 characters is 1 beat per character from the top When the search information is expressed in Kanji, there is a method of extracting one Kanji character as one beat.

  That is, the speech interactive search device 401 includes the search information database in which search information to be searched is stored, and beats extracted by disassembling the search information stored in the search information database. The first beat, which is the beat constituting the head of each search information, is stored in the order of frequency used in the search information database, and the beat following the first beat is also stored in the order of use frequency When the user inputs the search information by voice, the recognition processing means for preferentially recognizing the first beat in the beat database from the frequently used beat in the search information database, and the result of recognizing the first beat, If it is determined that it is a strong leading beat that is likely to be a beat input by the user, it will lead to the leading leading beat, and from the frequently used beat, Recognizing process first, extracting the dominant second beat, and outputting the recognition process repeating means for repeating the recognition process until the last beat constituting the search information, and when the recognition process for all beats is completed. If the total score satisfies a condition that allows the user to specify the search information only by performing a confirmation process, the total score is calculated for each search information. It is an example of a voice interactive information retrieval apparatus having a dialog means for determining information.

  In this case, as a result of recognition processing for the first beat and the second beat, if the total score does not satisfy the above-mentioned condition that the user can specify the search information only by performing confirmation processing, the second beat recognition result and the third beat During the integration of the recognition results, the remaining number of beats is defined as the number of objects that can be processed in real time in order of frequency of use, and has parallel processing means for parallel processing in order from the first beat. .

  In this case, the parallel processing means is means for performing parallel processing on the remaining objects within the question and answer time of the confirmation processing when the user performs confirmation processing. In addition, when the user does not perform the confirmation process, the number of beats to be selected as the recognition target is determined in consideration of the time required for the recognition process.

  Further, the voice interactive search device 401 has a search beat in which the first beat, which is the beat constituting the head of each search information, is extracted from the beats obtained by disassembling the search information stored in the search information database. The beats that are stored in the order of frequency used in the database and that follow the first beat are also prioritized from the most frequently used beats in the search information database for the first beat in the beat database that is stored in order of use frequency. As a result of recognizing and processing the first beat, if it is determined that the beat is likely to be a beat input by the user, it is connected to the leading beat and is frequently used. Recognizing processing is preferentially performed from beats, the second leading beat is extracted, and the recognition processing repeating means for repeating the recognition processing until the last beat constituting the search information, and all beats When the recognition process ends, the total score calculation means for calculating the total score for each output search information, and if the total score satisfies the condition for specifying the search information only by the user confirming process, It is an example of the voice interactive information search device which has the dialogue means which performs a necessary dialogue between and confirms the search information.

  Further, the beat database 482 stores the first beat, which is the beat located at the beginning of the search information, in the order of the highest frequency of use and the first beat after the first beat. For each of the second beat group in which the second beat, which is the connected beat, is arranged in order of frequency of use, and the second beat, the third beat, which is the beat connected next to the second beat, is the frequency of use. It is an example of the beat database which has the 3rd beat group arranged in descending order.

  If this is grasped as the invention of the method, the first beat, which is the beat located at the head of the search information, is arranged in order of frequency of use, and each of the head beats is connected next to the first beat. Arranging the second beat, which is a beat, in order of frequency of use, and for each of the second beats, arranging the third beat, the beat connected to the second beat, in order of frequency of use. This is a method for creating a beat database.

  The voice interactive search device 401 can be grasped as a program invention. That is, among the beats extracted by disassembling the search information stored in the search information database, the first beat that is the beat constituting the head of each search information is used in the search information database. A recognition process procedure for preferentially recognizing the first beat in the beat database stored in order of frequency of use and stored in order of frequency from the most frequently used beat in the search information database. As a result of recognizing the leading beat, if it is determined that the leading beat is likely to be input by the user, the leading beat is connected to the leading leading beat, and the recognition process is preferentially performed from the frequently used beat. Then, extracting the leading second beat and repeating the recognition process until the last beat constituting the search information, and the time when the recognition process for all beats is completed When the total score satisfies the total score calculation procedure for calculating the total score for each output search information and the condition that the user can specify the search information only by performing confirmation processing, the necessary dialogue is performed with the user. The interactive procedure for determining the search information can be grasped as a program for causing a computer to execute.

  In addition, the first beat, which is the beat located at the beginning of the search information, is arranged in order of frequency of use, and the second beat that is the beat connected to the first beat is used for each of the first beats. Understanding as a program that causes a computer to execute the procedure of arranging in order of frequency of use and the procedure of arranging the third beat, which is the next beat after the second beat, in order of frequency of use for each of the second beats. Can do.

Furthermore, the program may be recorded on a recording medium such as an FD, CD, MD, DVD, HD, optical disk, magneto-optical disk, or semiconductor memory.

It is a block diagram which shows the search apparatus 1 of the compound noun inputted by the speech which is the 1st Example of this invention. It is a figure which shows the specific example in which the corporate name in the search database 81 is comprised with a compound noun, and this compound noun is divided and registered for every single noun which comprises it. 5 is a diagram showing a specific example of a search assistance database 82 in the search device 1. FIG. 4 is a flowchart specifically illustrating a processing process for determining a search word in the search device 1. In the search device 1, it is a figure which shows the specific example of the recognition result list | wrist with which the recognition result with respect to each single noun which comprises the search word (compound noun) input by speech and the recognition likelihood correspond. It is a figure which shows the mode of an actual search word candidate list preparation in the search device. 6 is a diagram illustrating a specific example of an updated recognition result list in the search device 1. FIG. It is a figure which shows the actual search word candidate list obtained as a result of performing the process which produces a real search word candidate list in the search device. 2 is a diagram generally showing a search database 81 in the search device 1. FIG. It is a figure which shows the search assistance database 82 generally. It is a block diagram which shows the search apparatus 101 of the compound noun input by the speech which is the 2nd Example of this invention. It is a figure which shows the specific example in which the corporate name in the search database 181 is comprised with a compound noun, and this compound noun is divided and registered for every single noun which comprises it. In a 2nd Example, it is a figure which shows the specific example of the search assistance database 182. FIG. In the 2nd example, it is a flow chart which shows the processing process which finalizes a search word concretely. In a 2nd Example, it is a figure which shows the specific example of the recognition result list | wrist with which the recognition result with respect to each single noun which comprises the search word (compound noun) input by speech and the recognition likelihood correspond. In a 2nd Example, it is a figure which shows the specific example of the updated recognition result list | wrist. In a 2nd Example, it is a figure which shows the mode of a real search word candidate list creation. It is a block diagram which shows the search apparatus 201 of the compound noun inputted by the speech which is the 3rd Example of this invention. In the compound noun search apparatus 201, it is a figure which shows the specific example in which the corporate name in the search database 281 is comprised with a compound noun, and this compound noun is divided and registered for every single noun which comprises it. It is a figure which shows the specific example of the order search auxiliary database 282 in the search device 201 of a compound noun. It is a figure which shows the specific example of the reverse order search assistance database 283 in the compound noun search apparatus 201. FIG. 10 is a flowchart showing a process for determining a search word in the compound noun search device 201. It is a figure which shows the specific example of the recognition result list at the time of searching in order in the compound noun search apparatus 201. FIG. It is a figure which shows the specific example of the recognition result list at the time of searching in reverse order in the search apparatus 201 of a compound noun. In the compound noun search apparatus 201, it is a figure which shows the specific example of the recognition result list | wrist when the recognition result at the time of an order search, and the recognition result at the time of a reverse order search are merged. In the compound noun search apparatus 201, it is a figure which shows the state just before creation of a real search word candidate list. It is a flowchart which shows the operation | movement of the 4th Example of this invention. It is a block diagram which shows the search apparatus 301 of the voice interactive compound noun which is the 5th Example of this invention. In the search apparatus 301, it is a figure which shows the specific example in which the corporate name in the search database 381 is comprised with a compound noun, and this compound noun is divided and registered for every single noun which comprises it. 5 is a diagram showing a specific example of data stored in a search auxiliary database 382 in the search device 301. FIG. 5 is a flowchart specifically illustrating a processing process for determining a search word in the search device 301. In the search device 301, it is a figure which shows the specific example of the recognition result list | wrist with which the recognition result with respect to each single noun which comprises the search word (compound noun) input by speech and the recognition likelihood correspond. In the search device 301, it is a figure which shows the example of a partial matching search word candidate, a partial matching likelihood, and the example of a re-recognition result. It is a block diagram which shows the voice interactive type | formula interactive search apparatus 401 which is the voice interactive type search apparatus 401 which is the 6th Example of this invention. It is a figure which shows an example of the whole image of the search information database 481 used with the speech interactive search device 401. FIG. It is a figure which shows an example of the beat database 482 used with the speech interactive search device 401. FIG. It is a figure which shows the specific example of the search information database in the voice interactive search device. 7 is a diagram showing an example of beats stored in a beat database 482 in the voice interactive search device 401. FIG. It is a flowchart which shows the operation | movement which determines the search information in the speech interactive search device 401. FIG. In the speech interactive search device 401, the first beat candidates in the beat database 482 corresponding to the first beat constituting the search information input by the user by voice are arranged in order of frequency of use (in descending score), and the arranged heads are arranged. It is a figure which preferentially recognizes the predetermined number of first beats among beat candidates and shows the recognition result. It is a figure which shows the integration result (total score) of the score which is the recognition result of a leading beat candidate, and the score which is the recognition result of a 2nd beat candidate in the speech interactive search device 401. FIG.

Explanation of symbols

1 ... Search device for compound nouns input by voice,
2 ... Voice input part,
3 ... voice recognition unit,
4 ... voice recognition result output unit,
5 ... Real search word candidate list creation part,
6 ... Dialogue control unit,
7 ... Audio output unit,
8 ... System database,
81 ... Search database,
82 ... Search auxiliary database,
83 ... YES / NO database,
G1 ... first single noun set group,
G2: second single noun set group,
G3: Third set of nouns,
G4 ... Fourth single noun set group,
101 ... Compound noun search device inputted by voice,
102 ... voice input unit,
103 ... voice recognition unit,
104 ... voice recognition result output unit,
105 ... Real search word candidate list creation unit,
106 ... Dialog control unit,
107: Audio output unit,
108 ... system database,
181 ... Search database,
182 ... Auxiliary search database,
183 ... YES / NO database,
Ga1 ... first noun set group,
Ga1-1 ... the first sub-set group in the first-stated single noun set group Ga1,
Ga1-2 ... the second sub-set group in the first-stated single noun set group Ga1,
Ga1-3 ... the third sub-set group in the first-stated single noun set group Ga1,
Ga2 ... the single noun group of the second notation,
Ga2-1 ... the first sub-set group in the second-stated single noun set group Ga2,
Ga2-2 ... the second sub-set group in the second-stated single noun set group Ga2,
Ga2-3 ... the third sub-set group in the second-stated single noun set group Ga2,
201 ... Search device for compound nouns input by voice,
202 ... voice input unit,
203 ... voice recognition unit,
203S ... voice recognition device,
204 ... voice recognition result output unit,
204a ... voice recognition result list creation unit,
205 ... Real search word candidate list creation unit,
206 ... Dialog controller,
207 ... Audio output unit,
207S ... audio output device,
208 ... System database,
281 ... Search database,
282 ... Order search auxiliary database,
283 ... Reverse search auxiliary database,
284 ... YES / NO database,
301 ... Voice interactive compound noun search device,
302 ... voice input unit,
303 ... voice recognition unit,
303S ... voice recognition software,
304 ... voice recognition result output unit,
305 ... Real search word candidate list creation unit,
306 ... Dialog control unit,
307: Audio output unit,
307S ... Audio output software,
308 ... System database,
381 ... Search database,
382 ... Search auxiliary database,
383 ... YES / NO database,
Gc1... First noun set group,
Gc1-1 ... the first sub-set group in the first-stated single noun set group Gc1,
Gc1-2 ... the second sub-set group in the first-stated single noun set group Gc1,
Gc2: the second-stated single noun set group,
Gc2-1 ... the first sub-set group in the second-stated single noun set group Gc2,
Gc2-2: a second sub-set group in the second-stated single noun set group Gc2,
401 ... voice interactive search device,
402: Voice input unit,
403 ... voice identification unit,
431 ... voice recognition unit,
432 ... Recognition result adjustment unit,
404 ... Dialog control unit,
405 ... Audio output unit,
405S: Audio output software,
406 ... voice recognition device,
408 ... system database,
481 ... Search information database,
482 ... Beat database.

Claims (10)

A search information database storing search information that is character information to be searched;
Of the beats extracted by disassembling the search information stored in the search information database, the first beat that is the beat constituting the head of each search information is used in the search information database. A beat database stored in order of frequency of use and also stored in order of frequency of use for the beat following the first beat;
Recognition processing means for preferentially recognizing the first beat in the beat database from the frequently used beat in the search information database when the user inputs the search information by voice;
As a result of recognizing the first beat, the first beat whose recognition score of the recognized first beat exceeds the first threshold is determined to be a leading first beat, and is a beat that follows the leading first beat and is used Recognition processing repeating means for performing recognition processing preferentially from high beats, extracting a leading second beat, and repeating the recognition processing until the last beat constituting the search information;
Total score calculating means for calculating a total score of the score of the first beat and the score of the second beat for each search information output at the time when recognition processing for all beats is completed;
When the total score exceeds the second threshold value, the search word candidate in the real search word candidate list is displayed on a predetermined display, and a dialog is executed with the user to confirm the search information. Interactive means to do;
A voice interactive information retrieval apparatus comprising: In claim 1,
If the total score does not satisfy the above-mentioned conditions that allow the user to specify the search information only by performing confirmation processing, the confirmation processing is performed sequentially from the first beat, and confirmation of the search information is realized by performing the confirmation for each beat. A voice interactive information retrieval apparatus characterized by that. In claim 1,
As a result of the recognition processing for the first beat and the second beat, if the total score does not satisfy the above-mentioned condition that the user can specify the search information only by performing confirmation processing, the recognition result of the second beat and the recognition of the third beat While integrating the results, for each remaining beat, set the number that can be processed in real time in the order of frequency of use as the recognition target, and having parallel processing means that performs parallel processing in order from the first beat A featured voice interactive information retrieval device. In claim 3,
The parallel processing means is a means for performing parallel processing on the remaining objects within a question and answer time of the confirmation processing when the user performs confirmation processing. In claim 3,
A voice interactive information retrieval apparatus, wherein when a user does not perform confirmation processing, each beat number to be selected as a recognition target is determined according to a time required for recognition processing. In claim 1,
The method of extracting the beat by decomposing the search information is a method of extracting one hiragana character as one beat when the search information is expressed in hiragana (however, stuttering, sound repellent, and long sound are combined with the previous hiragana one beat. Counting),
When the above search information is written in hiragana, it is cut out in order from the beginning for each predetermined hiragana character number, with 1 beat (however, stuttering, sound repelling, and long sound are counted as one character together with the previous hiragana),
When the search information is written in hiragana, according to the number of constituents of the written hiragana, how to cut out the number of characters constituting one beat,
When the search information is expressed in Kanji, how to extract one Kanji character as one beat,
Among them, a voice interactive information search device characterized in that any one of them is cut out. In claim 6,
When the search information is written in hiragana, the cutout method for defining the number of characters constituting one beat in accordance with the number of hiragana written in the search is that search information consisting of 4 characters is 1 every 2 characters from the beginning. A voice interactive information search apparatus characterized in that the search information consisting of three beats is cut out by counting three letters from the beginning by counting one beat for each character. Of the beats extracted by disassembling search information that is character information stored in the search information database, the first beat that is the beat constituting the top of each search information is used in the search information database. A recognition process for preferentially recognizing the first beat in the beat database stored in order of use frequency from the most frequently used beat in the search information database for the beats that are stored in the order of frequency of use and that follow the first beat. With means;
As a result of recognizing the first beat, the first beat whose recognition score of the recognized first beat exceeds the first threshold is determined to be a leading first beat, and is a beat that follows the leading first beat and is used Recognition processing repeating means for performing recognition processing preferentially from high beats, extracting a leading second beat, and repeating the recognition processing until the last beat constituting the search information;
Total score calculating means for calculating a total score of the score of the first beat and the score of the second beat for each search information output at the time when recognition processing for all beats is completed;
When the total score exceeds the second threshold, the search word candidate in the real search word candidate list is displayed on a predetermined display, thereby performing a dialogue with the user, and A means of dialogue to establish;
A voice interactive information retrieval apparatus comprising: Of the beats extracted by disassembling search information that is character information stored in the search information database, the first beat that is the beat constituting the top of each search information is used in the search information database. For the first beat in the beat database stored in order of use frequency, the voice recognition unit preferentially selects the beat that is frequently used in the search information database. A recognition processing stage for performing speech recognition processing;
As a result of recognizing the first beat, the first beat whose recognition score of the recognized first beat exceeds the first threshold is determined to be a leading first beat, and is a beat that follows the leading first beat and is used A recognition processing repetition stage in which recognition processing is preferentially performed from beats having a high value, a leading second beat is extracted, and the recognition processing is repeated until the last beat constituting the search information;
A total score calculation stage for calculating a total score of the score of the first beat and the score of the second beat for each search information output when the recognition processing for all the beats is completed;
When the total score exceeds the second threshold value, the search word candidate in the real search word candidate list is displayed on a predetermined display, and a dialog is executed with the user to confirm the search information. A dialogue stage to do;
A voice interactive information retrieval method characterized by comprising: Of the beats extracted by disassembling search information that is character information stored in the search information database, the first beat that is the beat constituting the top of each search information is used in the search information database. For the first beat in the beat database stored in order of use frequency, the voice recognition unit preferentially selects the beat that is frequently used in the search information database. Recognition processing procedure for voice recognition processing;
As a result of recognizing the first beat, the first beat whose recognition score of the recognized first beat exceeds the first threshold is determined to be a leading first beat, and is a beat that follows the leading first beat and is used A recognition processing repetition procedure that performs recognition processing preferentially from high beats, extracts the dominant second beat, and repeats the recognition processing until the last beat constituting the search information;
A total score calculation procedure for calculating the total score of the score of the first beat and the score of the second beat for each output search information when the recognition processing for all beats is completed;
When the total score exceeds the second threshold value, the search word candidate in the real search word candidate list is displayed on a predetermined display, and a dialog is executed with the user to confirm the search information. Interactive steps to do;
A program that causes a computer to execute.
(more than)

Images