JPH0713591A - Device and method for speech recognition - Google Patents

Abstract

PURPOSE:To improve the recognition performance in noisy environment even when a large vocabulary are registered by using all recognized object vocabulary which are previously recognized as objects of recognition in less noisy environment, and excluding some of the recognized vocabulary which are previously registered from objects of recognition and performing recognition in the noisy environment. CONSTITUTION:This device and method are equipped with a speech input part 101, an A/D conversion part 102 which quantizes an input speech obtained by the speech input part 101, an analysis part 103 which obtains feature components of the input speech, a standard pattern storage part 104 for feature vectors of the previously registered recognized object vocabulary a collation part 105 which recognizes the input speech by obtaining the similarity degree between the feature vectors and the feature vectors obtained by the analysis part 103, a noise state detection part 106, and a recognized object vocabulary limiting part 107 which makes the number of the recognized object vocabulary to be collated by the collation part 105 less than the number of the previously registered recognized object vocabulary. The recognized vocabulary limiting part 107 limits the recognized object vocabulary corresponding to the noise state detected by the noise detection part 106.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device,
In particular, the present invention relates to a portable voice recognition device that operates stably even under different noise conditions depending on the environment.

[0002]

2. Description of the Related Art In a small portable terminal represented by a portable telephone, the number and size of operation buttons are limited due to its small size. Therefore, such a device is desired to be operated by using handwritten characters or voice rather than being operated by buttons.

A problem of voice recognition is that when used in a noisy environment, the recognition performance is significantly deteriorated due to ambient noise. Moreover, in the case of recognition of a large vocabulary or recognition of many similar words, a recognition error is likely to occur. Therefore, in order to put the voice recognition device into practical use, a noise resistant technology that can correctly recognize even a voice uttered under noise is indispensable.

[0004]

As means for correctly recognizing a voice uttered in a noisy environment, a method of removing noise from a voice on which noise is superimposed by preprocessing or a voice on which noise is superimposed can be correctly recognized. A recognition method is needed. For the former, the one that removes noise using an adaptive filter, the noise spectrum mixed in the speech is estimated,
There is a spectral subtraction method that subtracts from the input spectrum. The latter includes a method that uses parameters and distance measures that are less susceptible to noise, and a noise superposition method that preliminarily superimposes noise on a standard pattern. However, although many noise processing methods have been proposed, it is still insufficient when compared with the recognition performance in a quiet environment.

An object of the present invention is to provide a voice recognition device capable of improving the voice recognition accuracy in a noisy environment.

[0006]

In order to solve the above-mentioned problems, a voice recognition apparatus according to the present invention includes a voice input section for inputting a voice to be recognized, and an input voice obtained from the voice input section. A / D conversion unit for converting the input speech, an analysis unit for obtaining a characteristic component of the input voice, a standard pattern storage unit storing a feature vector of a recognition target vocabulary registered in advance, and a standard pattern storage unit stored in the standard pattern. A matching unit that recognizes the input voice by finding the similarity between the feature vector and the feature vector found by the analysis unit, a noise state detection unit that detects the noise state at the time of recognition, and the matching unit collate. A recognition vocabulary limiting unit for reducing the number of recognition target vocabularies to a number smaller than the number of recognition target vocabularies registered in advance, wherein the recognition vocabulary limiting unit recognizes the recognition target vocabulary according to the noise state detected by the noise state detecting unit. Vocabulary In which it was to be limited.

Further, the voice recognition method according to the present invention is a voice recognition method in a voice recognition device for performing voice recognition by collating a plurality of vocabularies registered in advance with an input voice,
Detecting the surrounding noise state at the time of voice recognition, determining the magnitude of the detected noise, when it is determined that the noise is small,
Speech recognition is performed using all the registered vocabulary as recognition targets, and when it is determined that the noise is large, only a part of the registered vocabulary is used as recognition target for speech recognition. And

[0008]

Function: A recognizable vocabulary is registered in advance in the voice recognition device. From the standpoint of the user, it is desirable that the number of vocabularies that can be registered be as large as possible. However, in general, a speech recognition apparatus has lower recognition performance as the vocabulary to be recognized is larger and the similar words are larger. Especially, in a noisy environment, the degree of performance deterioration due to similar words is large. Therefore, if a large number of vocabularies that can be registered are set, sufficient recognition performance can be obtained in a noise-free environment, but in a noise environment, recognition errors are large and sufficient recognition performance cannot be expected.

Therefore, in the present invention, when used in an environment with little noise, all the recognition vocabulary registered in advance is used as a recognition target, and in a noisy environment, the noise state detecting section and the recognition vocabulary limiting section are used. The recognition is performed by excluding a part of the recognition vocabulary registered in advance (for example, removing a vocabulary or a similar word that is rarely used from the recognition target).

By the above operation, it becomes possible to improve the recognition performance in a noisy environment even when many vocabularies are registered.

[0011]

EXAMPLES Examples of the present invention will be described below. In the present embodiment, an example in which the present invention is applied to a mobile terminal device having a voice recognition function will be described.

FIG. 15 shows the external appearance of the portable terminal device.
In the figure, 101 is a voice input unit, 110 is a voice output unit, 111 is a display unit, 112 is a selection button, and 506 is a sensor unit. Functions of these units will be described later.

FIG. 1 is a block diagram showing an embodiment of the system configuration of the speech recognition apparatus of the present invention. In FIG. 1, 101 is a voice input unit, 102 is an A / D conversion unit, 10
3 is an analysis unit, 104 is a standard pattern storage unit, 105 is a collation unit, 106 is a noise state detection unit, 107 is a recognition vocabulary restriction unit, 108 is a voice section detection unit, and 109 is a command execution unit. The voice input unit 101 is a unit for inputting voice such as voice commands. The voice signal input from the voice input unit 101 is quantized by the A / D conversion unit 102, and the voice section is detected by the voice section detection unit 108. For more information on voice segment detection, see Furui "Digital Speech Processing" Tokai University Press. For example, a method is often used in which the short-time power of the input signal is extracted at regular time intervals and the voice section is determined based on whether or not the short-time power equal to or more than a threshold value is continued for a predetermined time or longer. It is almost certain that the portion other than the voice section detected by the voice section detection unit 201 is considered to be a noise-only section. The quantized voice signal in the voice section enters the analysis unit 103. The analysis unit 103 extracts an audio feature vector used for recognition determination by using an analysis method such as LPC analysis. For details on voice feature extraction methods, see Furui "Digital Speech Processing" mentioned above. As the voice feature vector, for example, LPC cepstrum is often used. Standard pattern storage unit 104
Stores a vocabulary feature vector (standard pattern) registered in advance. The matching unit 105 calculates the similarity between the input speech converted into the feature vector by the analysis unit 103 and the standard pattern stored in the standard pattern storage unit 104, and recognizes the registered vocabulary having the largest similarity. Output as a result. The noise state detection unit 106 is a unit that detects a noise state, which is one of the major factors that deteriorate the recognition performance. The recognition vocabulary limiting unit 107 limits the recognition target vocabulary according to the degree of the noise state obtained by the noise state detecting unit 106. Of course, when the noise level is sufficiently low, the recognition vocabulary limiting unit 107 does nothing.

The system processing flow of the system shown in FIG. 1 will be described with reference to FIG. First, when the user activates a voice command, the A / D conversion unit 102 is activated and voice acquisition is started (S10). Therefore, the voice section of the input voice signal is detected (S11) and the noise state is detected (S12). This noise state is compared with a predetermined threshold value (S13), and if it is determined that the noise level is low, step S
Proceed to 15. If it is determined that the noise is large, the number of recognition target words is limited (S14). Regarding the limitation of the recognition target vocabulary, for example, the feature vector (standard pattern) of the registered vocabulary stored in the standard pattern storage unit 104 is provided with the usage frequency information, and only the frequently used one is set as the recognition target vocabulary. Therefore, the standard pattern of the restricted vocabulary is compared with the input voice (S15), the recognition result is presented (S16), and the user's confirmation is requested. This presentation is shown in Figure 1.
The display unit 111 or the voice output unit (including the voice synthesis unit) 110 shown in FIG. When the user inputs that the recognition result is correct by pressing the confirmation button or the like, the command represented by the recognition result is executed (S18). If the recognition result is incorrect, the processing from step S10 is performed again by the user giving a voice instruction again. When the user indicates that the recognition result is correct, the usage frequency information is further updated (S19).

The input / output sequence of this system will be described with reference to FIG. 13 by taking a voice dial as an example. First,
When the voice command is activated, guidance for requesting the command is output from the system (2001). In this example, the output from the system indicates a voice "Please input a command.", But the voice may be displayed on the screen by liquid crystal or the like, or both may be used in combination. The user inputs a voice command according to the guidance (2002). In this example, the name "Ichikawa" of the person who wants to dial is uttered. When a voice command is input, the system returns the recognition result of the input voice command to the user (2003) and asks the user for confirmation. Here, an example is shown in which "Ichikawa" is mistakenly recognized as "Ishikawa". When the user determines that the recognition result is incorrect, the user re-inputs the voice command (“Ichikawa”) (2004). The system recognizes the voice command in which the voice command is re-input, and returns the recognition result (“Ichikawa” in this example) (2005). Therefore, the user notifies the system that the recognition result is correct by pressing the confirmation button or the like (2006). as a result,
The system executes the recognized command and starts dialing (2007).

Next, the noise state detecting section 106 will be described in detail. FIG. 2 shows a configuration example of the noise state detection unit 106.

In FIG. 2, reference numeral 202 is a power calculation unit. The power calculation unit 202 receives a signal other than the voice section from the voice section detection unit 201 as a noise section signal, and calculates the short-time power of the signal to obtain the power of ambient noise. The obtained ambient noise power is output to the recognition vocabulary limiting unit 107 as noise state information at the time of recognition. Of course, the power calculation unit 202 is not necessary when the short-term power is calculated by the voice section detection unit 108.

FIG. 3 is a diagram for explaining another example of the configuration of the noise state detecting section 106.

In FIG. 3, 302 and 302 'are power calculation units having the same configuration as the power calculation unit 202 in FIG.
Is a voice-to-noise ratio (S / N) calculator. As in the case of FIG. 2, the input voice quantized by the A / D conversion unit 102 is separated by the voice period detection unit 108 into a period in which voice exists and a period other than that. Power calculator 302,
302 'calculates the short-term power of the voice section signal and the short-term power of the noise section signal, respectively. Of course, when the short-term power is calculated in the voice section detection unit 108, the power calculation units 302 and 302 'are not necessary. S /
The N calculation unit 303 calculates the S / N of the input voice based on the short-time power of the voice section signal and the short-time power of the noise section signal.
Is output to the recognition vocabulary limiting unit 107 as noise state information at the time of recognition.

FIG. 4 shows a system configuration of another embodiment of the present invention having a noise state detecting section 408 having still another configuration.

In FIG. 4, as in the embodiment of FIG.
1 is a voice input unit, 102 is an A / D conversion unit, 103 is an analysis unit, 104 is a standard pattern storage unit, 105 is a collation unit, 40
6 is a recognition result confirmation unit, 107 is a recognition vocabulary restriction unit, 108
Is a voice section detection unit, 109 is a command execution unit, and 408 is a noise state detection unit. Similar to the embodiment of FIG. 1, the audio signal input from the audio input unit 101 is A /
The D conversion unit 102 quantizes and the analysis unit 103 converts it into a feature vector. The matching unit 105 calculates the distance between the feature vector output from the analysis unit 103 and the standard pattern stored in the standard pattern storage unit 104, and outputs the recognition result to the recognition result confirmation unit 406. The recognition result confirmation unit 406 outputs the result recognized by the collation unit 105 as information such as voice and characters, and requests the user to confirm the recognition result. The confirmation method of the recognition result is "Confirm", "Cancel"
The operation of the selection button 112 (FIG. 15) such as, confirmation by voice, rewording, etc. can be considered. Recognition result confirmation unit 40
When confirming that the recognition result is correct, 6 outputs the recognition result to the command execution unit 109 and requests the execution of the voice command. Further, when the recognition result confirmation unit 406 determines that the recognition result is not correct, the recognition result confirmation unit 406 outputs to the noise state detection unit 408 that a recognition error has occurred, and receives a new recognition result. When detecting that a recognition error has occurred, the noise state detection unit 408 outputs the frequency of the recognition error to the recognition target restriction unit 107 as noise state information. Recognition vocabulary restriction unit 107
Limits the recognition target vocabulary according to the degree of the noise state obtained by the noise state detection unit 408.

FIG. 5 is a diagram for explaining the third embodiment of the noise state detecting section 106.

In FIG. 5, 101 is a voice input unit and 10 is a voice input unit.
2 is an A / D conversion unit, 103 is an analysis unit, 104 is a standard pattern storage unit, 105 is a collation unit, 107 is a recognition vocabulary restriction unit,
Reference numeral 108 denotes a voice section detection unit, 506 denotes a sensor unit, 5
Reference numeral 07 is a noise state detection unit. Similar to the embodiment of FIG.
The audio signal input from the audio input unit 101 is quantized by the A / D conversion unit 102 and converted into a feature vector by the analysis unit 103. The matching unit 105 calculates the distance between the feature vector output from the analysis unit 103 and the standard pattern stored in the standard pattern storage unit 104, and obtains a recognition result. The sensor unit 506 measures, for example, the distance between the recognition device and the user's (mouth), and outputs it to the noise state detection unit 507. The distance can be measured by using a distance sensor such as an infrared sensor or an ultrasonic sensor. The noise state detection unit 507 assumes that the S / N of the input voice is worse as the distance between the voice input unit 101 and the user is larger,
The distance information obtained by the sensor unit 506 is output to the recognition target restriction unit 107 as noise state information. The recognition vocabulary limiting unit 107 limits the recognition target vocabulary according to the degree of the noise state obtained by the noise state detecting unit 507. Here, a switch for selecting a recognition condition is provided in the recognition device, and the state of the switch is detected instead of measuring the distance by the sensor unit 506, and the noise state detection unit 507 detects noise state information at the time of recognizing the state of the switch. It is also possible to realize a modification in which the output is output to the recognition vocabulary restriction unit 107 as. The switch may be a “normal mode”, a “noise mode”, or the like, or may be a switch that can be switched in multiple stages for each usage environment.

The configuration for detecting the noise state described above is merely an example, and other configurations may be used as long as the same effect can be obtained. It is also possible to use a plurality of the above-mentioned configurations together.

Next, the recognition vocabulary limiting unit 107 will be described in detail. FIG. 6 shows a configuration example of the recognized vocabulary restriction unit 107.

In FIG. 6, reference numeral 104 is a standard pattern storage unit, 602 is an important vocabulary storage unit, and 603 is a recognized vocabulary restriction unit. The important vocabulary storage unit 602 is the standard pattern storage unit 1
Of the registered vocabulary stored in 04, the vocabulary frequently used by the user and a previously designated important vocabulary are stored. The important vocabulary may be specified either during vocabulary registration or after registration. If the value of the noise information input from the noise state detection unit 106 is larger than a threshold value, the recognition vocabulary restriction unit 603 stores in the important vocabulary storage unit 602 among the registered vocabulary stored in the standard pattern storage unit 104. Matching unit 1 with only registered vocabulary as recognition target vocabulary
Output to 05. According to this configuration example, the vocabulary designated by the user is not excluded from the recognition target even if the recognition target vocabulary is limited in a noisy environment, and there is a concern that the important vocabulary cannot be recognized in the noise. There is no need.

As another configuration example of the recognition vocabulary limiting unit 107, the number of recognition target vocabularies is determined according to the recognized noise state, and the vocabulary registered in the standard pattern storage unit 104 is restricted from the one having the highest priority. A method of selecting as a vocabulary will be described. Here, the case where the noise state detection unit 106 outputs the S / N value of the input voice will be described, but the same can be considered when the noise state detection unit 106 is other cases.

As a means for realizing such a configuration, FIG. 7 shows a recognition target vocabulary number restriction table 700 used in the recognition vocabulary restriction unit 107 for limiting the number of recognition target words. In FIG. 7, 701 is the S / N value of the voice input from the noise state detection unit 106, and 702 is the target S.
Indicates the number of recognition target vocabulary when / N. Recognition vocabulary restriction unit 107
Can refer to the table 700 of FIG. 7 to obtain the number of recognition target vocabularies according to the S / N input from the noise state detection unit 106. Recognition vocabulary restriction unit 107
Selects the standard patterns for the number of vocabularies in descending order of priority from the standard pattern storage unit 104 according to the number of recognition target vocabularies.

As described above, the method of selecting the recognition target vocabulary in descending order of priority is to add the usage frequency information to each registered vocabulary stored in the standard pattern storage unit 104, and to use the vocabulary in descending order of usage frequency. It can be realized by selecting as a recognition target vocabulary. Thus, by using the usage frequency information to limit the recognition target vocabulary, it is possible to significantly reduce the probability that the input voice deviates from the recognition target vocabulary.

As a third configuration example of the recognized vocabulary restriction unit 107, a method of giving similar word information to each registered vocabulary stored in the standard pattern storage unit 104 will be described.
As a means for realizing this method, a similar word table 8 of the registered vocabulary stored in the standard pattern storage unit 104 is shown in FIG.
Indicates 00.

In FIG. 8, reference numeral 801 is a registered vocabulary, and 802.
Indicates a similar registered vocabulary for each registered vocabulary, and 803 indicates the similarity between the registered vocabulary and the similar registered vocabulary. In this configuration example,
When a new vocabulary is registered in the standard pattern storage unit 104,
The similarity is calculated between the vocabulary that was originally registered and the newly registered vocabulary, and the registered vocabulary with the highest similarity is registered as a similar registered vocabulary in the similar word table 800 together with the similarity at that time. . The recognition vocabulary restriction unit 107
By referring to the similar word table 800, one of the registered vocabulary pairs having a similarity of a certain value or more is excluded from the recognition target.
At this time, as the similarity threshold, a correspondence table between the output of the noise state detection unit 106 and the similarity threshold is prepared in advance as in FIG. 7. Further, the determination of which vocabulary in the similar word pair to select as the recognition target can be realized by giving priority to a vocabulary that is frequently used or a vocabulary registered by the user as an important vocabulary.

According to the configuration example of FIG. 8, when used in an environment with little noise, all the recognition vocabularies registered in advance are used as recognition targets, and in a noisy environment, the noise state detection unit and the recognition vocabulary are used. The restriction unit recognizes a vocabulary and a similar word having a low frequency of use among the recognition vocabularies registered in advance by excluding them from the recognition target. As a result, even when many vocabularies are registered, it is possible to improve the recognition performance in a noisy environment. Further, except for the case where the selection switch described with reference to FIG. 5 is provided, the recognition target vocabulary is automatically limited according to the noise state, so that there is an advantage that the user does not have to be aware of the usage environment.

In the embodiment described above, it is impossible to recognize a vocabulary which is not recognized because of the limitation of the recognition target vocabulary. Therefore, as a modification of the present embodiment, a method of replacing the recognition target vocabulary when the recognition target vocabulary is not correctly recognized as a result of being restricted will be described. The method of confirming the recognized result has already been described in the recognition result confirmation unit 406 of FIG.

FIG. 9 schematically shows a set of recognition target vocabulary when recognition is performed twice. In FIG. 9, 901 is a set of all registered vocabularies, 902 is a set of first recognition target vocabularies selected by the recognition vocabulary limiting unit 107, and 903 is a set of second recognition target vocabularies.

As shown in the processing flow of FIG. 14, in the first recognition (S20) under noise, recognition is performed on the vocabulary group 902 (S21, S23). Therefore, if the input vocabulary does not exist in the recognition target vocabulary group 902 (S24, S25), a correct recognition result cannot be obtained. Therefore, when the first recognition is not correct, the recognition vocabulary restriction unit 107 restricts the recognition target vocabulary again for the vocabulary excluding the first recognition target vocabulary 902 from the entire registered vocabulary 901 (S22).
Note that the example of FIG. 9 shows a case in which the number of vocabularies obtained by excluding the set 902 of the first-time recognition target vocabularies from the total registered vocabularies 901 is larger than the number of recognition vocabularies restricted by the recognition vocabulary restriction unit 107. Conversely, it may be smaller than the number of recognized vocabularies limited by the recognized vocabulary limiting unit 107. In this case, it is possible to use all of the vocabularies excluding the set 902 of the first-time recognition target vocabulary from the entire registered vocabulary 901 for the second recognition, and to use the first recognition-target vocabulary. It is possible to make a modification such that the frequently used vocabulary and the vocabulary registered as the important vocabulary are not excluded as the second recognition target vocabulary. According to the modification of the above-described embodiment, it is possible to correctly recognize a vocabulary that is not a recognition target due to the restriction of the recognition target vocabulary in the second and subsequent recognitions. It should be noted that it is conceivable to remove the limitation on the number of vocabularies and make all the vocabularies the recognition target at the second time.

Next, as a second example for recognizing a vocabulary that is not a recognition target due to the limitation of the recognition target vocabulary,
A fourth embodiment of the present invention will be described in which a plurality of recognition results are presented when the recognition target vocabulary is limited using similar word information. FIG. 10 is a system configuration diagram for explaining the present embodiment, and FIG. 11 is a schematic diagram of recognition target vocabulary.

In FIG. 10, 101 is a voice input unit, 1
Reference numeral 02 is an A / D conversion unit, 103 is an analysis unit, 104 is a standard pattern storage unit, 105 is a collation unit, 106 is a noise state detection unit, 108 is a voice section detection unit, 1007 is a recognition vocabulary restriction unit, and 800 is a similar word. A table, 1009 is a recognition result confirmation unit, and 109 is a command execution unit. As in the embodiment of FIGS. 1 and 4, the audio signal input from the audio input unit 101 is quantized by the A / D conversion unit 102 and converted into a feature vector by the analysis unit 103. The collating unit 105 is the analyzing unit 10.
3 and the feature pattern output from the standard pattern storage unit 10
The distance from the standard pattern stored in 4 is calculated. At this time, the recognition vocabulary limiting unit 1009 limits the recognition target vocabulary based on the information in the similar word table 800, according to the noise state detected by the noise state detecting unit 106. As described in FIG. 8, the similar word table 800 stores the registered vocabulary and the similar vocabulary in pairs. The recognition result confirmation unit 1009 outputs the result recognized by the collation unit 105 as information such as voice and characters, and requests the user to confirm the recognition result.

In FIG. 11, 1101 is a set of all registered vocabularies, 1102 is an example of similar word pairs stored in the similar word table 800 of FIG. 10, 1103 is a recognition target vocabulary selected by the recognition restriction unit 1007. Is a set of. In the example of FIG. 11, a pair of similar words “Ichikawa”,
Since "Ichikawa" of "Ishikawa" is the recognition target vocabulary, "Ichikawa" is recognized by the matching unit 105 even when the input from the voice input unit 101 is "Ishikawa". Therefore, the recognition confirming unit 1009 displays both the vocabulary (“Ichikawa”) recognized by the collating unit 105 as a recognition candidate and the vocabulary of the similar pair (“Ishikawa”).
1 (FIG. 15) or the voice output unit 110 (FIG. 15) to present the user with recognition candidates. At this time, in order to emphasize the difference between the two, the non-identical parts "chi" and "shi" may be emphasized and presented. The recognition candidate can be selected by using a button, a touch panel, or the like, or can be input by voice again. Further, the recognition confirmation unit 100
A modification of 9 is possible such that, when the similarity of the similar word pair of the vocabulary recognized by the matching unit 105 is low, the similar word is not output as a recognition candidate. When the recognition result confirmation section 1009 confirms the recognition result, the recognition result is sent to the command execution section 109.
And request execution of the voice command. Here, it goes without saying that each component may use any implementation method of the above-described embodiments.

According to the embodiment of FIG. 10 described above, when the recognition target vocabulary is limited by using the similar word information, a plurality of recognition results are presented, so that the recognition target vocabulary is excluded from the recognition target vocabulary. It is also possible to make a candidate for recognition even for a vocabulary that has been recognized.

[0040]

According to the present invention, all the recognition target vocabularies registered in advance are used as recognition targets when used in a low noise environment, and registered in advance in a noisy environment. By recognizing a part of the recognized vocabulary by excluding it from the recognition target, it is possible to improve the recognition performance in a noisy environment even when many vocabularies are registered.

[Brief description of drawings]

FIG. 1 is a block diagram showing a system configuration of an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration example of a noise state detection unit in FIG.

FIG. 3 is a block diagram showing another configuration example of the noise state detection unit in FIG.

FIG. 4 is a block diagram showing a system configuration of a second exemplary embodiment of the present invention.

FIG. 5 is a block diagram showing a system configuration of a third exemplary embodiment of the present invention.

6 is a block diagram showing a configuration example of a recognition target vocabulary number limiting unit in FIG. 1. FIG.

FIG. 7 is an explanatory diagram of a recognition target vocabulary number restriction table showing another configuration example of the recognition target vocabulary number restriction unit of FIG. 1.

8 is an explanatory diagram of a similar word table that can be used in the embodiment of FIG.

9 is an explanatory diagram of a recognition target vocabulary in the example of FIG. 1 and the like.

FIG. 10 is a block diagram showing a system configuration of a fourth exemplary embodiment of the present invention.

11 is an explanatory diagram for explaining a recognition target vocabulary in the embodiment of FIG.

FIG. 12 is a flowchart showing a system process of the embodiment of FIG.

13 is an explanatory diagram of an input / output sequence example in the embodiment of FIG.

14 is a flowchart of system processing corresponding to the description of FIG.

FIG. 15 is an external view of a portable terminal device to which the present invention is applied.

[Explanation of symbols]

101 ... Voice input section, 102 ... A / D conversion section, 103 ...
Analyzing unit, 104 ... Standard pattern storing unit, 105 ... Collating unit,
106 ... Noise state detecting unit, 107 ... Recognition vocabulary limiting unit, 1
08 ... voice section detection unit, 109 ... command execution unit

Claims (17)

[Claims] 1. A voice input unit for inputting a voice to be recognized, and an A / D for quantizing the input voice obtained from the voice input unit.
A conversion unit, an analysis unit that obtains the characteristic component of the input speech, a standard pattern storage unit that stores a characteristic vector of a recognition target vocabulary that is registered in advance, a characteristic vector that is stored in the standard pattern, and the analysis The matching unit that recognizes the input voice by finding the similarity to the feature vector obtained by the unit, the noise state detection unit that detects the noise state at the time of recognition, and the number of recognition target vocabularies that the matching unit collates. A recognition vocabulary limiting unit that reduces the number of recognition target vocabularies registered in advance, wherein the recognition vocabulary limiting unit limits the recognition target vocabulary according to the noise state detected by the noise state detecting unit. A voice recognition device. 2. The noise state detection unit detects a noise level when the voice recognition device is used.
The voice recognition device described. 3. The voice recognition device according to claim 1, wherein the noise state detection unit detects the noise state by detecting the voice-to-noise ratio of the voice input from the voice input unit. 4. The noise state detection unit detects the distance between the voice input unit and the mouth of the user, and detects the noise state based on the detected distance. Voice recognition device. 5. A condition selection unit for a user to select a usage condition, and the noise condition detection unit detects a noise condition based on the condition of the condition selection unit.
The voice recognition device described. 6. An erroneous recognition determination section for determining that the voice input from the voice input section is not correctly recognized, wherein the noise state detection section is based on the state of the erroneous recognition determination section. The voice recognition device according to claim 1, wherein 7. A designated vocabulary storage unit for storing a vocabulary designated by a user from among the vocabularies registered in the standard pattern storage unit, and the recognized vocabulary restriction unit is stored in the designated vocabulary storage unit. 7. The voice recognition device according to claim 1, wherein the recognition target vocabulary is limited based on the vocabulary information that is stored. 8. A similar vocabulary storage unit that stores a relationship of similar vocabulary with respect to a vocabulary registered in the standard pattern storage unit, and the recognized vocabulary restriction unit is similar vocabulary information of the similar vocabulary storage unit. 7. The speech recognition apparatus according to claim 1, wherein the recognition target vocabulary is limited based on the above. 9. A use frequency storage unit for storing a past use frequency for a vocabulary registered in the standard pattern storage unit, and the recognition vocabulary restriction unit stores frequency information in the use frequency storage unit. 7. The voice recognition device according to claim 1, wherein the recognition target vocabulary is limited based on the original. 10. A similar vocabulary storage unit for storing a relationship of similar vocabulary with respect to a vocabulary registered in the standard pattern storage unit, and a past use frequency for a vocabulary registered in the standard pattern storage unit. The recognition vocabulary limiting unit limits the recognition target vocabulary based on the similar vocabulary information of the similar vocabulary storage unit and the frequency information of the vocabulary frequency storage unit. The voice recognition device according to claim 1. 11. An erroneous recognition determination unit that determines that the voice input from the voice input unit is not correctly recognized, and the recognition result of recognition in which the recognition target vocabulary is restricted by the recognition vocabulary restriction unit is incorrect. 11. If there is, the restriction of the recognition target vocabulary part is removed, and all the vocabulary registered in the standard pattern storage part is recognized as a recognition target, and the speech according to claim 1. Recognition device. 12. An erroneous recognition determination unit that determines that the voice input from the voice input unit is not correctly recognized, and the recognition result of recognition in which the recognition target vocabulary is restricted by the recognition vocabulary restriction unit is incorrect. If there is, the remaining vocabulary excluding the recognition target vocabulary selected in the recognition target vocabulary unit out of all the vocabularies registered in the standard pattern storage unit is recognized as a recognition target. 1 to 10
The voice recognition device according to any one of 1. 13. An erroneous recognition determination unit that determines that the voice input from the voice input unit is not correctly recognized, and the recognition result of recognition in which the recognition target vocabulary is restricted by the recognition vocabulary restriction unit is incorrect. If there is, a part of the remaining vocabulary excluding the recognition target vocabulary selected in the recognition target vocabulary part out of all the vocabularies registered in the standard pattern storage part is recognized as a recognition target. The voice recognition device according to any one of claims 1 to 10. 14. A presentation unit for presenting a recognition result to a user, wherein when the recognition vocabulary limiting unit performs recognition with the recognition vocabulary limited, the presentation unit, together with the recognized vocabulary, 11. The voice recognition device according to claim 8, wherein a similar word of the recognition vocabulary stored in the similar vocabulary storage unit is presented as a recognition candidate. 15. When the recognition vocabulary is restricted by the recognition vocabulary limiting unit, the recognition unit stores the recognition result stored in the similar vocabulary storage unit. 11. The speech recognition apparatus according to claim 8, wherein the presentation unit presents the recognized vocabulary and the similar word as recognition candidates only when the degree of similarity of the vocabulary to the similar word is high. 16. The presenting unit, when presenting the recognized vocabulary and the similar word to the presenting unit, emphasizes and presents a non-identical portion of the presented vocabulary. 15. The voice recognition device according to 15. 17. A voice recognition method in a voice recognition device for performing voice recognition by collating a plurality of pre-registered vocabularies with an input voice, wherein a surrounding noise state at the time of voice recognition is detected, and the detection is performed. When it is determined that the noise is small, the speech recognition is performed using all the registered vocabularies as recognition targets, and when it is determined that the noise is large, the registration is performed. A speech recognition method characterized by performing speech recognition by using only a part of the vocabulary as a recognition target.