JP2003330491A - Method, device, and program for voice recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device which enables a user to easily specify a cause of erroneous recognition. <P>SOLUTION: The voice recognition device has a voice input part 1 for outputting voice data corresponding to an input voice, a voice recognition part 2 for collating the voice data with standard patterns related to preliminarily registered voice recognition vocabulary to recognize the input voice, a condition detection part 3 for examining whether at lease one of a plurality of different factors causing trouble in voice recognition exists in the voice data or not, a voice recognition result discrimination part 4 for judging that the voice recognition is proper when any of the factors is not detected by the condition detection part 3 and judging that the voice recognition is not proper when some of the factors are detected, and an operation part 5 for executing one of a plurality of preliminarily set response operations corresponding to the factors respectively in the case that it is judged that the voice recognition is not proper. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus and method for recognizing speech in a real environment, assuming that ambient noise or ambient noise such as ambient conversation is input as unnecessary speech. Furthermore, the present invention relates to a program for executing such real environment-oriented voice recognition.

[0002]

2. Description of the Related Art A voice recognition device for recognizing a voice of a speaker by acoustic analysis (spectral analysis) is known. This speech recognition device includes a limited vocabulary speech recognition device that inputs a limited vocabulary (predetermined vocabulary) and an arbitrary vocabulary speech recognition device that inputs a vocabulary of an arbitrary vocabulary. Some of them have been put into practical use. . There are several types of limited vocabulary speech recognizers, but among them, many discrete products have been provided so far because the discrete word speech recognizer that recognizes the voice divided into words is technically easy. ing.

The main part of the discrete word voice recognition device is a memory unit in which a short-time spectrum relating to a voice recognition vocabulary (limited vocabulary) is registered in advance as a standard pattern, and a speaker's voice (input voice) input from a microphone. For example, the acoustic analysis unit that obtains the short-time spectrum in frame units of 10 to 20 ms and the matching between the short-time spectrum obtained by the acoustic analysis unit and the standard pattern registered in the memory unit are examined, and the most matched standard pattern Is output as a recognition result.

As the matching process with the standard pattern performed in the voice recognition section, for example, dynamic programming (DP: Dyna) is used.
A DP matching method utilizing a mic programming method is known. In this DP matching method, a standard pattern corresponding to a candidate word to be recognized is created in advance, and matching is performed while keeping the time axis correspondence between the feature amount pattern obtained by analyzing the input voice and all the standard patterns. By doing so, the most similar standard pattern is extracted.

A problem may occur in voice recognition due to unnecessary voice such as ambient noise. In order to prevent the occurrence of the problem of voice recognition due to this unnecessary voice, Japanese Patent Laid-Open No. 9-146586 discloses a voice that prompts a speaker to speak loudly when ambient noise is large. A recognizer is disclosed. In this voice recognition device, it is determined whether or not the energy of the input voice is smaller than a predetermined threshold value, and when it is determined that the energy is low, the user is warned that a loud voice is uttered.

Further, in voice recognition, there is a problem of insufficient voice volume of the speaker. As a means for solving this problem, Japanese Unexamined Patent Application Publication 2000-2000
Japanese Patent Publication No. 155600 discloses a device that warns a speaker to speak at an appropriate input voice level when the input voice level is not within a proper range. In addition, there is a method disclosed in Japanese Patent Laid-Open No. 2000-75893 that determines whether or not voice data is appropriate for voice recognition before performing voice recognition.

[0007]

In the speech recognition, when unnecessary noise such as environmental noise or surrounding conversation is superimposed on the speech of the speaker, the accuracy of the speech analysis is lowered and the recognition rate of the speech is lowered. Such unnecessary voice is not only received from the surrounding environment, but the system itself may be a noise source. Also, in a quiet situation with little ambient noise,
Depending on the environment, it may be affected by the echo sound.
On the other hand, as a general problem of speech recognition, recognition accuracy deterioration due to utterance defects such as insufficient utterance other than the recognition vocabulary and insufficient voice volume due to lack of knowledge of the user's utterance method, or speaking before the recognition engine operates. There are also problems.

As described above, in real-world voice recognition, there are many factors that hinder the voice recognition and they are complicated. Therefore, in the above-described conventional voice recognition device, it is difficult for the user to specify the cause of the erroneous recognition when the erroneous recognition occurs. Further, when the ambient noise is erroneously recognized in the voice recognition vocabulary, the voice recognition device malfunctions unexpectedly as a result, but it is difficult for the user to specify the cause of the malfunction.

In the method disclosed in Japanese Unexamined Patent Publication No. 9-146586, when the energy of the input voice is smaller than a predetermined threshold value, the user is warned that a loud voice is issued. It is not designed to identify multiple factors that interfere with voice recognition.

Even in the one disclosed in Japanese Unexamined Patent Publication No. 2000-155600, when the input voice level is not within the proper range, only a warning is given to the speaker to speak at the proper input voice level, After all, it is not designed to identify a plurality of factors that hinder the above-mentioned voice recognition.

Also in Japanese Unexamined Patent Publication No. 2000-75893, it is only necessary to determine whether or not voice data is appropriate for performing voice recognition before performing voice recognition. It is not designed to identify the factors.

An object of the present invention is to provide a voice recognition device, a voice recognition method, and a program which can solve the above problems and allow a user to easily identify the cause of erroneous recognition.

[0013]

In order to achieve the above object, a voice recognition device of the present invention comprises a voice input means for outputting voice data corresponding to an input voice, a voice recognition vocabulary registered in advance with the voice data. A voice recognition means for recognizing the input voice by collating with a standard pattern regarding whether or not there is at least one of a plurality of different factors causing a problem in the voice recognition in the voice data. When neither of the plurality of factors is detected by the situation detecting means and the situation detecting means, it is determined that the voice recognition is valid, and when any of the plurality of factors is detected. In, the voice recognition result determination means for determining that the voice recognition is not appropriate, and a plurality of response operations respectively corresponding to the plurality of factors are preset, When in serial speech recognition result determining means is determined that the speech recognition not reasonable,
A response unit that selectively executes a corresponding response operation from among the plurality of response operations.

In the voice recognition method of the present invention, the first step of outputting voice data corresponding to the input voice and the input voice by collating the voice data with a standard pattern relating to the voice recognition vocabulary registered in advance. Second voice recognition
And a third step of checking whether at least one of a plurality of different factors causing a problem in the voice recognition is present in the voice data, and the plurality of factors in the third step. If none of the above is detected, it is determined that the voice recognition is valid, and if any of the plurality of factors is detected, it is determined that the voice recognition is not valid. If the voice recognition is determined not to be appropriate in step 4 and the fourth step, it corresponds to one of a plurality of preset response operations corresponding to the plurality of factors. A fifth step of selectively performing a response operation.

The program of the present invention verifies the input voice by collating the first process for outputting the voice data corresponding to the input voice with the standard pattern relating to the voice data and the voice recognition vocabulary registered in advance. A second process of recognizing, a third process of checking whether or not at least one of a plurality of different factors causing a problem in the voice recognition exists in the voice data, and a third process. If none of the plurality of factors is detected, it is determined that the voice recognition is valid, and if any of the plurality of factors is detected, the voice recognition is not valid. And a plurality of responses corresponding to the plurality of factors set in advance when the voice recognition is determined to be invalid in the fourth process. Characterized in that to execute a fifth process for performing responding operation appropriate from among the operations selectively to the computer.

In the present invention as described above, when any one of a plurality of different factors that cause a problem in voice recognition is included in the voice data, the factor is detected. Specifically, various noises such as stationary noises, sudden noises, and reverberant sounds (including noises of the system itself) and unnecessary voices such as surrounding conversations, as well as various problems that interfere with voice recognition such as utterance defects. The factor is detected. Then, the response action corresponding to the detected factor is appropriately selected from the response actions prepared in advance and executed, so that the user can easily determine the cause of the misrecognition from the response action. Is possible.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings.

(Embodiment 1) FIG. 1 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a first embodiment of the present invention. This voice recognition device includes a voice input unit 1, a voice recognition unit 2, a situation detection unit 3, a voice recognition result determination unit 4, an operation unit 5, and a voice recognition vocabulary storage unit 21.

The voice input unit 1 is provided with a microphone, for example, and when a voice uttered by a speaker (spoken voice) is input from the microphone, voice data corresponding to the input voice (converted into a frequency domain at predetermined time intervals) Voice pattern)
Is output. Speech recognition vocabulary storing section 21 is intended to store the vocabulary necessary for voice recognition (voice recognition vocabulary), for example, "Good morning", "Hello" word of standard patterns such as (voice pattern) in advance to speaker of voice Will be registered based on.

The voice recognition unit 2 includes a voice recognition vocabulary storage unit 21.
It has a well-known voice recognition engine that uses the voice recognition vocabulary stored in the dictionary as a dictionary, and performs voice recognition of the voice data output from the voice input unit 1. Specifically, the voice recognition unit 2
Cuts out the voice section of the voice data output from the voice input unit 1 and checks the data of the cut voice section for matching with the voice recognition vocabulary (standard pattern) stored in the voice recognition vocabulary storage unit 21, The most matched speech recognition vocabulary is output as the recognition result 2a. This recognition result 2a is supplied to the voice recognition result determination unit 4,
At the same time, the data of the voice section cut out at the time of the recognition (cutout voice data 2b) is supplied to the situation detection unit 3.

The situation detecting section 3 performs a spectrum analysis on the cut-out voice data 2b supplied from the voice recognizing section 2 to check whether or not the cut-out voice data 2b has a cause of a problem in terms of environment and voice utterance. , And outputs the result (situation detection result) to the voice recognition result determination unit 4. The cause of environmental problems is, for example, stationary noise that is continuous noise such as surrounding conversation and noise (wind noise of a fan). As a cause of the problem of voice utterance, there are utterance problems such as head cut, lack of power, and excess.

The voice recognition result judgment unit 4 judges whether or not there is a fatal defect in the voice recognition in the voice recognition unit 2 based on the situation detection result supplied from the situation detection unit 3, and a fatal malfunction is caused. If there is not, the recognition result 2a supplied from the voice recognition unit 2 is sent to the operation unit 5 as it is as valid, and if there is a fatal defect, the recognition result supplied from the voice recognition unit 2 is recognized. The result 2a is discarded and the content of the defect is sent to the operation unit 5 instead.

When the recognition result 2a is supplied from the voice recognition result judging unit 4, the operation unit 5 executes a preset response operation based on the recognition result 2a, and the voice recognition result judging unit 4 makes a fatal error. When the content of the specific malfunction is supplied, a response (avoidance) operation according to the content of the malfunction is executed. Here, the preset response operation is, for example, to respond to the recognition result 2a with a voice message prepared in advance. The avoidance operation means, for example, notifying the speaker of a warning message according to the content of the defect, or actively performing the operation of avoiding the problem in the voice recognition device. As an active defect avoidance operation, for example, when there is a lot of ambient noise, the volume of the microphone is turned down, and when recognition errors continue, voice recognition is stopped and other input devices such as mouse, touch panel, and buttons are activated. There is an operation to give guidance to use.

Next, the operation of the speech recognition apparatus of this embodiment will be specifically described.

First, when performing voice recognition, the user registers the voice pattern (standard pattern) of a word necessary for voice recognition with his or her voice in the voice recognition vocabulary storage unit 21. For example, "Hello", "Good evening", such as "Good morning", to register the speech patterns of a variety of words. A well-known method can be used for the registration of this voice pattern, and the procedure thereof is general, so a detailed description thereof will be omitted here.

After the voice word is registered in the voice recognition vocabulary storage unit 21, the user inputs a voice based on the registered word. The voice uttered by the user is input to the voice input unit 1, and the voice input unit 1 outputs voice data (voice pattern) corresponding to the input voice. At this time, if there is ambient noise (stationary noise such as a conversation with another person in the surroundings), the ambient noise is included in the voice data output from the voice input unit 1.

When the voice data output from the voice input unit 1 is supplied to the voice recognition unit 2, the voice recognition unit 2 cuts out a necessary voice section from the voice data, and outputs the cut-out voice section data (cutout). The voice data 2b) is checked for matching with the voice recognition vocabulary (standard pattern) stored in the voice recognition vocabulary storage unit 21. For this matching, for example, the DP matching method can be used. Then, the voice recognition unit 2 supplies the best-matched voice recognition vocabulary as the recognition result 2a to the voice recognition result determination unit 4, and also supplies the cut-out voice data 2b to the situation detection unit 3.

Speech data 2b cut out from the speech recognition unit 2
When the cutout voice data 2b is supplied, the situation detection unit 3 checks whether or not there is a cause of a defect due to ambient noise or poor voice utterance, and the result of the check (situation detection result) is determined as a voice recognition result. Supply to part 4. Since ambient noise and poor voice utterance exist as unique spectral patterns in the cut-out voice data 2b, spectrum analysis of the cut-out voice data 2b is performed to detect such a spectral pattern, which results from ambient noise and poor voice utterance. The cause of the failure can be detected.

When the situation detection result is supplied from the situation detection unit 3, the voice recognition result determination unit 4 determines whether or not there is a fatal defect in the voice recognition in the voice recognition unit 2 based on the situation detection result. Judge (judge validity). For example, the user utters a word other than the voice recognition vocabulary stored in the voice recognition vocabulary storage unit 21 due to insufficient recognition of the voice utterance method for performing normal voice recognition, or the amount of voice is insufficient. If this happens, it will result in poor voice production. If there is such a voice utterance defect during voice input in the voice input unit 1, a fatal defect occurs in voice recognition in the voice recognition unit 2, resulting in erroneous recognition. The voice recognition result determination unit 4 examines a fatal defect in the voice recognition in the voice recognition unit 2 due to such poor voice utterance from the situation detection result from the situation detection unit 3. Then, if there is no fatal defect, the voice recognition result determination unit 4 sends the recognition result 2a supplied from the voice recognition unit 2 to the operation unit 5 as it is, and the fatal defect exists. In this case, the recognition result 2a supplied from the voice recognition unit 2 is discarded, and instead the fault content is sent to the operation unit 5.

When the recognition result 2a is supplied from the voice recognition result judging section 4, the operating section 5 executes a preset operation as a response to the recognition result 2a. For example, when the user voice-inputs the word "Ohayo" in the voice input unit 1 and the voice recognition result determination unit 4 supplies the effective recognition result 2a "Ohayo", the operation unit 5
The response message "Good morning" is output. This response message can also be presented by voice.

On the other hand, when the fault content is supplied from the voice recognition result judging section 4, the operating section 5 is preset.
The response operation corresponding to the failure content is executed from the plurality of response operations corresponding to the plurality of factors. For example, if the content of the problem is insufficient voice volume, the user is instructed to speak in a louder voice. Also, when the problem content that the ambient noise is large is supplied, guidance is given to reduce the volume of the microphone. Such guidance is executed by, for example, appropriately selecting a corresponding message from a message group for guidance prepared in advance.

According to the voice recognition apparatus of the present embodiment described above, when a fatal defect occurs in voice recognition due to any of a plurality of factors such as ambient noise and poor voice utterance, the user is informed of the fact. Guidance is provided according to the content of the defect. Therefore, the user can easily identify the cause of misrecognition.

The above-described voice recognition device of the present embodiment is a specific speaker voice recognition in which voice recognition is performed based on the voice recognition vocabulary registered in the voice recognition vocabulary storage unit 21 by the user's own voice input. However, the present invention is not limited to this. For example, it may be configured to perform unspecified speaker voice recognition in which voice recognition is performed in consideration of variations in voice due to differences in speakers.

Further, the voice recognition in the voice recognition unit 2 is not limited to the discrete word voice recognition, and various kinds of voice recognition can be applied. However, in the case of using a rejection vocabulary dictionary, which will be described later, it is preferable to use discrete word speech recognition because of the restriction of the rejection vocabulary dictionary.

(Embodiment 2) FIG. 2 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a second embodiment of the present invention. This voice recognition device is provided with the configuration of the first embodiment described above (see FIG. 1) as a basic configuration, and in order to make a more detailed response at the time of recognition, the rejection vocabulary storage unit 2
2 is provided. 2, the same components as those shown in FIG. 1 are designated by the same reference numerals. Here, in order to simplify the description, description of the same operation is omitted,
Only the characteristic part will be described in detail.

The reject vocabulary storage unit 22 stores in advance a reject vocabulary for rejecting an input voice except when the user correctly utters the voice recognition vocabulary stored in the voice recognition vocabulary storage unit 21. . The rejected vocabulary is, for example, a combination of vowels and syllables such as "a" and "ii", and ideally, the rejected vocabulary is as much as possible from the voice patterns other than the voice recognition vocabulary stored in the voice recognition vocabulary storage unit 21. It is desirable to register as. However, when the amount of rejected vocabulary increases, the recognition process (rejection process) using the rejected vocabulary takes time. As described above, there is a trade-off relationship between the amount of rejected vocabulary and the processing time, and it is necessary to set the rejected vocabulary in consideration of this point.

To generate the rejected vocabulary, for example, a voice recognition vocabulary stored in the voice recognition vocabulary storage unit 21 and its similar vocabulary are selected from a huge rejection vocabulary database using a syllable net composed of arbitrary syllable strings. The rejected vocabulary that matches each is removed. For the generation of the similar vocabulary, for example, a method using acoustic similar word generation is known. This method picks up a pair that is likely to cause misrecognition at the phoneme level in advance, and based on syllables (similar syllables) that have the same vowel with the paired phoneme as a consonant, similar to the speech recognition vocabulary. Generate a word. For example, if the phonemes in the “Ka” and “Ha” lines are similar phonemes to the “kenkai (view)” registered as the speech recognition vocabulary, “kenkai”, “henkai”, “ken” Four similar words "Yes" and "Henye" are generated. By using the rejected vocabulary thus obtained as a word dictionary for rejection, it is possible to reject the input voice or noise other than when the user correctly utters the voice recognition vocabulary.

In the voice recognition apparatus of this embodiment, when voice data is input from the voice input unit 1, the voice recognition unit 2
A voice section is cut out from the input voice data, and it is checked whether or not the voice pattern of the cut-out voice data does not match any voice pattern of the reject vocabulary stored in the reject vocabulary storage unit 22.

When the voice pattern of the input voice data matches the reject vocabulary, the voice recognition unit 2 rejects the input voice and sends a voice recognition result indicating that to the voice recognition result determination unit 4. , And outputs the cut-out voice data to the situation detection unit 3. The situation detection unit 3 detects factors such as ambient noise and speech defect detection, and outputs the detection result (situation detection result) to the voice recognition result determination unit 4.

When the voice recognition result (rejection) is input from the voice recognition unit 2, the voice recognition result determination unit 4 determines the malfunction content based on the situation detection result from the situation detection unit 3 as the operation unit 5.
Output to. When the fault content is supplied from the voice recognition result determination unit 4, the operation unit 5 executes a response operation according to the fault content. For example, if the voice volume is insufficient, the user is instructed to speak a little louder, and if the ambient noise is large, the user is instructed to turn down the volume of the microphone. Further, if there is no utterance defect despite the utterance, guidance is given to stop voice recognition or request voice input again. For example, the rejected vocabulary storage unit 22 stores a voice pattern (standard pattern) of “A” as a rejected vocabulary, and the voice input unit 1 has no voice utterance defect or ambient noise, and indicates “A” as unnecessary speech. When the voice input is performed and the fault content indicating only the voice recognition result (rejection) is output from the voice recognition result determination unit 4 to the operation unit 5, the operation unit 5 makes a rejection response by asking "what". Any of the above guidances can be performed by, for example, presenting a message for guidance prepared in advance by voice.

On the other hand, when the input voice pattern does not match the rejected vocabulary, the voice recognition unit 2 determines that the input voice pattern is one of the voice patterns of the voice recognition vocabulary stored in the voice recognition vocabulary storage unit 21. Check if there is a match. After that, the operation described in the above-described first embodiment is performed.

According to the present embodiment described above, the input voice except when the user correctly utters the voice recognition vocabulary stored in the voice recognition vocabulary storage unit 21 is rejected, and the voice about the rejected voice is rejected. Since the recognition process is not performed, the recognition rate is higher than that in the above-described first embodiment.

Further, with respect to the rejected ones, the defect factor is detected, and the response operation is performed in accordance with the detected defect factor. Therefore, a finer response than that of the first embodiment described above. Is possible.

(Third Embodiment) FIG. 3 is a block diagram showing a schematic configuration of a speech recognition apparatus according to a third embodiment of the present invention. In the voice recognition device, a history analysis unit 6 is added in the configuration of the second embodiment described above (see FIG. 2) in order to make a finer response at the time of erroneous recognition. Figure 3
The same parts as those shown in FIG. 2 are designated by the same reference numerals. Here, in order to simplify the description, description of the same operation will be omitted, and only the characteristic part will be described in detail.

The history analysis unit 6 holds as a history the voice recognition result output from the voice recognition unit 2 and the situation detection result output from the situation detection unit 3 in the past, and the voice recognition result from the voice recognition unit 2 is retained. Is output and the situation detection result is output from the situation detection unit 3, the past occurrence states of the voice recognition result and the situation detection result are checked based on the history, and the checked occurrence state satisfies a predetermined condition. It is determined whether or not it is determined, and the determination result (history analysis result) is output to the voice recognition result determination unit 4 together with the voice recognition result. Here, the predetermined condition is the past continuity and frequency of the same situation detection result regarding the same voice recognition result. The continuity indicates how many times the same situation detection result is continuously output, and the frequency indicates how many times the same situation detection result is output in a predetermined number of times in the past.

The voice recognition result judging section 4 judges the presence or absence of excessive noise from the history analysis result based on the continuity and frequency inputted from the history analyzing section 6, and outputs the result to the operating section 5. The operation unit 5 performs a response operation prepared in advance according to the presence or absence of excessive noise from the voice recognition result determination unit 4. For example, in the case of excessive noise, the voice recognition is stopped, and in the case of not excessive noise, the corresponding response operation is appropriately selected and executed from the preset response operation group.

When the situation detection unit 3 does not detect any environmental trouble (miscellaneous surroundings, etc.) and it is an utterance and there is an utterance defect, the history analysis unit 6 analyzes the history information. Instead, the voice recognition result from the voice recognition unit 2 and the situation detection result from the situation detection unit 3 may be directly passed to the voice recognition result determination unit 4. in this case,
The voice recognition result determination unit 4 and the operation unit 5 perform the same operations as in the above-described first and second embodiments.

A specific example of how to use the history analysis result will be described later.

(Fourth Embodiment) FIG. 4 is a block diagram showing a schematic structure of a speech recognition apparatus according to a fourth embodiment of the present invention. In this voice recognition device, in the configuration of the above-described third embodiment (see FIG. 3), the information detection unit 3 includes the environment estimation unit 31 in order to make a finer response at the time of erroneous recognition. There is. 4, those parts which are the same as those shown in FIG. 3 are designated by the same reference numerals. Here, in order to simplify the description, description of the same operation will be omitted, and only the characteristic part will be described in detail.

The environment estimating unit 31 estimates a defect due to noise or reverberant sound included in the clipped voice data input from the voice recognizing unit 2 and emitted from the surrounding environment and the system, and outputs the estimation result. Specifically, the environment estimation unit 31 detects the failure factors included in the cut-out voice data in the priority order of stationary noise, sudden noise, and reverberant sound. Here, the stationary noise means noise in which electrical noise, surrounding conversation, and surrounding environment noise continue continuously. On the other hand, the sudden noise is a sound that is suddenly emitted, such as a sound when an object is dropped or a sound when a hand is struck once, and has no continuity. The reverberant sound is a sound that is echoed on a wall inside the room and is superimposed on the uttered sound. In addition,
In the present embodiment, the priority order is set so as to give priority to the one having a large influence on the voice recognition when there is a defect factor, but all three defect factors of stationary noise, sudden noise, and echo sound are determined. However, all detection results may be output.

Next, the operation of the environment estimating unit 31 will be specifically described. FIG. 5 is a flow chart for explaining the operation of the environment estimation unit 31 shown in FIG.

First, the cut-out voice data is acquired (step S1), and it is determined whether or not the acquired cut-out voice data contains stationary noise (step S1).
2). If there is stationary noise, output “With stationary noise”
(Step S3), the process ends. If there is no stationary noise, it is subsequently determined whether or not the cutout voice data acquired in step S1 includes sudden noise (step S4). If there is a sudden noise, "there is a sudden noise" is output (step S5), and the process is terminated. If there is no sudden noise, it is subsequently determined whether or not the cut-out audio data acquired in step S1 includes reverberation (step S6). If there is a reverberant sound, "reverberant sound is present" is output (step S7), and the process ends.
If there is no reverberant sound, no detection result is output (step S8), and the process ends.

Next, a concrete description will be given of a method of detecting each of the defect factors such as stationary noise, sudden noise, and reverberation noise.

(Detection of Stationary Noise) The following two methods can be considered as a method of detecting stationary noise.

The first method is to pay attention to the voice length of the clipped voice data. In the presence of stationary noise, the speech recognition engine cannot accurately detect the end of the speech section. In this case, the cut-out voice section length (cut-out voice data length) is the one whose end is accurately detected, for example, the cut-out voice section length (cut-out voice data length when the voice recognition vocabulary is uttered in a quiet environment. ) Will be longer than. Therefore, ambient noise can be detected by acquiring the cut-out voice data length (time) and determining whether the value (time) exceeds a certain threshold value.

The certain threshold value depends on the maximum number of syllables in the speech recognition vocabulary and needs to be adjusted accordingly. Further, the cut-out voice section length (cut-out voice data length) can be calculated from the data size of the cut-out voice when the cut-out voice is dropped into data of a general voice format such as Wave format. For the start and end of the voice section, for example, the average power of the white noise section with no voice is learned, and the start point is the point when it exceeds a predetermined threshold based on that average power, and the point when it returns to the average power of the white noise section again. Can be detected as the end.

The second method is to pay attention to the voice power other than the utterance portion of the cut-out voice data. Short section before and after the cut-out audio data, for example, 100 ms before and after
The presence of stationary noise can be detected by checking whether or not the maximum power in the 200 ms section exceeds a certain threshold. The threshold of the voice power needs to be adjusted in consideration of the voice power of the ambient noise measured in the actual environment.

Since the first and second stationary noise detection methods described above have advantages and disadvantages depending on the quality of the voice to be detected, it is desirable to use them together. FIG. 6 is a flowchart showing a stationary noise detection procedure when the first and second stationary noise detection methods are used in combination.

First, the voice length of the cut-out voice data is acquired (step S10), and it is checked whether the obtained voice length exceeds a preset threshold value (step S).
11). If it exceeds the threshold, "with stationary noise" is output (step S12), and the process ends. If it does not exceed the threshold value, subsequently, the peak power in the section of 100 ms before and after the cut-out audio data is acquired (step S13), and it is checked whether the acquired front and rear peak powers both exceed the preset threshold value. (Step S14). If both the front and rear peak powers exceed the threshold value, "with stationary noise" is output (step S15), and the process ends. If either one of the front and rear peak powers does not exceed the threshold value, "no stationary noise" is output (step S16), and the process ends.

(Detection of Sudden Noise) In order to detect sudden noise and human speech separately, a method of detecting noise in a short section without continuity is effective. Although various methods are conceivable as the detection method, as a simple example, a method of dividing the cut-out voice data into frames having a fixed frame length and detecting sudden noise will be described below.

FIG. 7 is a flow chart showing one procedure of sudden noise detection. First, the voice length (time) of the cut-out voice data is acquired (step S20), the obtained voice length (time) is divided into predetermined frame lengths, numbers are assigned from 0 to each frame, and the divided voice data is divided. The total number of frames is calculated (step S21). Then, after initializing the counter (step S22), the peak power in the frame of the number corresponding to the counter value is calculated (step S23), and it is determined whether or not the peak power exceeds a certain threshold value (step S24). ).

If it is determined in step S24 that the threshold power is exceeded, the peak powers of the preceding and following frames are acquired (step S25), and the acquired peak powers are both the previous threshold (the threshold used in step S24). ) Is not exceeded (step S26).
If it is determined in step S26 that the threshold value is exceeded, "with sudden noise" is output (step S2).
7), the process ends.

If it is determined in steps S24 and S26 that the threshold value is not exceeded, the counter value is incremented by 1 (step S28), and the counter value is changed to step S21.
It is determined whether the number of frames is smaller than the number of frames calculated in step S29. If the counter is smaller than the number of frames, the process returns to step S23. When the counter reaches the number of frames, "no sudden noise" is output (step S30), and the process ends.

FIG. 8 is a waveform diagram showing an example of cut-out audio data. The horizontal axis represents time and the vertical axis represents voice power. In this example, there is a sudden noise at the beginning of the voice data (immediately after the start of voice input), and the uttered voice is behind it. In FIG. 8, a section surrounded by a dotted line is one frame sectioned by a certain frame length (time width).

In step S21, the cut-out audio data is divided into a plurality of frames as shown in FIG. 8, and the number of frames is calculated. For example, when the voice length of the cut-out voice data is 1 s and the frame length is 50 ms, the number of frames is 20.

The frames before and after the step S25 are specifically the following frames.
For example, if the number of frames is 20, 0, 1, and
When the numbers are assigned to 2, ... 18, 19 and the current counter value is 5, the process for the frame to which the number “5” is assigned is performed in step S24, and the number “5” is assigned in step S25. The frames to which "4" and "6" are assigned are processed as the preceding and following frames, respectively. The frame length can be accurately detected by considering the characteristics of sudden noise and the duration of human utterance.
It is necessary to set the optimum value so that human speech is not erroneously detected.

Further, the fixed threshold value in step S24 depends on the peak power of the utterance related to the voice recognition vocabulary, but ideally, it is desirable to consider the influence of the peak power of the sudden noise on the voice recognition.

(Detection of Reverberant Sound) In the configuration shown in FIG. 4, it is difficult to detect the reverberant sound when the voice input unit 1 is composed of a single microphone. Voice input unit 1
As the above, it is possible to detect the presence of a reverberant sound by using two microphones having different directions and calculating the difference between the sounds input to the respective microphones. Specifically, by changing the directions of the two microphones,
One microphone outputs audio data that does not contain much reverberant sound, and the other microphone outputs audio data that contains reverberant sound.By calculating the difference between the output of these microphones, the reverberant sound Can be detected. Then, when the detected level of the reverberant sound exceeds a predetermined threshold value, it is determined as the reverberant sound that causes a defect in the voice recognition.

The estimation result of the environment estimation unit 31 is supplied to the history analysis unit 6. Then, the history analysis unit 6, the voice recognition result determination unit 4, and the operation unit 5 appropriately perform the operations described in the above-described first to third embodiments.

(Fifth Embodiment) FIG. 9 is a block diagram showing the schematic arrangement of a speech recognition apparatus according to the fifth embodiment of the present invention. In this voice recognition device, in the configuration of the above-described fourth embodiment (see FIG. 4), the situation detection unit 3 further includes a speech detection unit 32 in order to make a finer response at the time of erroneous recognition. ing. 9, those parts which are the same as those shown in FIG. 4 are designated by the same reference numerals. Here, in order to simplify the description, description of the same operation will be omitted, and only the characteristic part will be described in detail.

The utterance detection unit 32 determines whether the cut-out voice data input from the voice recognition unit 2 includes a human utterance, and outputs the determination result. As a method of detecting the presence or absence of human utterance from cut-out voice data, there is a method of detecting by paying attention to the harmonic structure of the spectrum of human uttered voice (Reference: “Spectrum harmonic structure for speech recognition. Utilization ”Proceedings of the 1997 Autumn Meeting of the Acoustical Society of Japan, pp3-4). It is possible to separate human speech and noise by extracting and using the feature amount of this harmonic structure from the voice data.

The speech detection result of the speech detection section 32 is supplied to the history analysis section 6. Then, the history analysis unit 6, the voice recognition result determination unit 4, and the operation unit 5 appropriately execute the operations described in the above-described first to fourth embodiments.

According to the present embodiment, by adding the utterance detection unit 32, when the speech recognition result is the vocabulary of the speech recognition vocabulary storage unit 21 without the utterance detection,
It is possible to discard the voice recognition result. This can reduce false recognition and malfunction. When the speech recognition result is the vocabulary in the rejected vocabulary storage unit 22 and the utterance is detected, the operation unit 5 is caused to present the guidance that “the utterance was detected but could not be recognized”. Is possible. Thereby, a finer response can be provided.

Further, the discard of the voice recognition result based on the speech detection result of the speech detection unit 32 and the unnecessary speech rejection using the rejection word storage unit 22 described in the above second embodiment interpolate with each other. This has the effect of increasing the rejection accuracy. For example, when the rejected vocabulary is not used, the utterance detection unit erroneously detects that there is utterance in a situation where there is a surrounding conversation. In this case, the word corresponding to the surrounding conversation is registered in the reject vocabulary storage unit 22 as a reject vocabulary in advance, and the surrounding conversation is recognized as a reject vocabulary and rejected, so that the utterance detection unit performs It is possible to prevent a malfunction by falsely detecting a conversation with utterance. Further, as the inverse interpolation, when noise extremely close to the speech recognition vocabulary is input and the rejection by the rejection vocabulary does not go well, it is possible to discard the noise by detecting no utterance by the utterance detection unit 32. Is.

(Sixth Embodiment) FIG. 10 is a block diagram showing the schematic arrangement of a speech recognition apparatus according to the sixth embodiment of the present invention. In this voice recognition device, in the configuration of the fifth embodiment described above (see FIG. 9), the situation detection unit 3 uses the utterance defect detection unit 3 in order to make a finer response at the time of erroneous recognition.
3 is further included. In FIG. 10, those similar to those shown in FIG. 9 are designated by the same reference numerals. Here, in order to simplify the description, description of the same operation will be omitted, and only the characteristic part will be described in detail.

The speech deficiency detection unit 33 determines whether the cut-out voice data input from the voice recognition unit 2 satisfies a predetermined fault condition. This utterance defect detection unit 33
2 differs from the environment estimation unit 31 in that it detects not the environmental factors but the defect factors on the human side who speaks a voice. Although there are various causes of the utterance defect detected by the utterance defect detection unit 33, here, representative examples include head loss, insufficient power, and excessive power.

Among the above-mentioned trouble factors, the power shortage can be detected by whether or not the peak power of the whole cut-out voice is smaller than a certain threshold value, and the excessive power is whether the peak power is larger than a certain threshold value or not. Can be detected with. The remaining head loss detection is performed as follows by focusing on the head portion of the clipped voice.

FIG. 11 shows an example of a simple algorithm for detecting head loss. In this example, for the sake of simplification, the section of 100 ms at the beginning of the audio data is focused, but the section of interest is not limited to 100 ms.
If the following head loss detection is possible, the section of interest is 50
It may be ms or 30 ms.

First, the first 100 m of the cut-out audio data
The peak power in the s frame is acquired (step S40). Next, it is determined whether or not the acquired peak power exceeds a threshold value (step S41), and if it exceeds the threshold value, it is subsequently determined whether or not a speech sound is included in the first 100 ms frame ( Step S42).

If it is determined in step S41 that the threshold is not exceeded, and if it is determined in step S42 that no speech is included, "no head cut" is output (step S43). The process ends. If it is determined in step S42 that the speech is included, "head cut" is output (step S4).
4), the process ends.

As described above, the speech detection unit 32 detects the speech defect, and the detection result is supplied to the history analysis unit 6. Then, the history analysis unit 6 and the voice recognition result determination unit 4
In the operation unit 5, the operations described in the above first to fifth embodiments are appropriately executed.

FIG. 12 shows the environment estimating unit 31 and the speech detecting unit 3.
2. A specific example of the detection result of the utterance defect detection unit 33 and the response operation according to the detection result will be described. Hereinafter, a specific operation will be described with reference to FIG.

[0083] In the example shown in FIG. 12, "Hello" as a voice recognition vocabulary, "A" respectively as rejectable vocabulary speech recognition vocabulary storing section 21 is stored in the rejectable vocabulary storage unit 22, a single audio input On the other hand, one of the results must be returned. For simplification, the environment estimation unit 31 detects only stationary noise. Speech defect detection unit 3
In 3, it is assumed that any one of the detection results is output in the priority order of "head cut", "power shortage", and "excessive power". The history analysis unit 4 gives “condition: 3 consecutive” when the same situation detection result is repeated three times, and 3 when the result of the utterance detection unit is “uttered” and the voice recognition result is a rejected vocabulary. In the case of continuous consecutive times, “discard: 3 consecutive times” is returned as the history collation result. Further, if the stationary noise is continuous or if the erroneous recognition is continuous without the stationary noise, the speech recognition is stopped. In addition, in the table shown in FIG. 12, "-" is entered indicates "regardless of the result", and "NOT (A)" indicates that the detection result of A is not detected.

When the environment estimation unit 31 detects that "stationary noise is present", the user is informed of the presence of stationary noise regardless of other results including the voice recognition result. This is because it is difficult to correctly recognize the speech uttered in the environment with "steady noise", and assuming that there are multiple speech recognition vocabulary, even if recognized in the speech recognition vocabulary, it may be erroneous recognition. Because it is expensive. Further, when the situation of “with stationary noise” continues three times, the history analysis unit 4 outputs “Condition: 3 continuous”, and the operation unit 5 stops the voice recognition. When this voice recognition is stopped, if another input device such as a mouse, a touch panel, or a button is provided, guidance for promoting the use of the other input device may be given.

The result of the environment estimation unit 31 is “no detection result”.
If it is, the output of the speech detection unit 32 is referred to. If there is "utterance", some response operation is performed by referring to the output result of the speech failure detection unit 33, and if "no speech", no operation is performed and it is ignored. Incidentally, in the case of "no speech", if the detection accuracy of the speech detection section 32 is not sufficient, for example, when the speech recognition vocabulary "Hi" is output, it may discard the detection result.

In the case of "with utterance" described above, if any utterance defect is detected by the utterance defect detection unit 33, guidance for notifying the utterance defect is given. When no speech defect is detected, the operation differs depending on the voice recognition result. If the voice recognition result of the speech recognition vocabulary "Hello", a response operation corresponding to the "Hello". When the rejection vocabulary is "a", the normal rejection response operation such as "what" is asked, but when the utterance failure is not detected three times in a row, the history analysis unit 4 determines "discard: 3".
"Continuous" is output, and the operation unit 5 stops the voice recognition.

[0087] Incidentally, with respect to the detection result of the speech fault detector 33, similar to the speech detection section 32, in accordance with the deterioration degree of the speech recognition accuracy for the speech to be detected and speech defect, speech recognition result voice recognition vocabulary "Hello If it is ", the detection result may be discarded.

The voice recognition and response operations in each of the embodiments described above can be basically realized by a program, and existing computer systems (basic configuration is a storage device, an input device such as a keyboard or a mouse). , A control device that receives input from a display device such as a CRT or LCD and an input device to access a storage device, and controls the operation of the output device and the display device).
Can be run on. This program may be installed in advance in the hard disk of the computer, or may be provided by a recording medium such as a CD-ROM. Furthermore, this program may be provided (downloaded) via the Internet.

[0089]

As described above, according to the present invention,
The user can easily determine the environmental or artificial cause of erroneous recognition and eliminate the cause to obtain an ideal environment suitable for voice recognition and an ideal input voice. A voice recognition device having a higher recognition rate can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of a voice recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a schematic configuration of a voice recognition device according to a second embodiment of the present invention.

FIG. 3 is a block diagram showing a schematic configuration of a voice recognition device according to a third embodiment of the present invention.

FIG. 4 is a block diagram showing a schematic configuration of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 5 is a flow chart diagram for explaining the operation of the environment estimation unit shown in FIG.

FIG. 6 is a flowchart showing one detection procedure of stationary noise.

FIG. 7 is a flowchart showing one detection procedure of sudden noise.

FIG. 8 is a waveform diagram showing an example of clipped audio data.

FIG. 9 is a block diagram showing a schematic configuration of a voice recognition device which is a fifth embodiment of the present invention.

FIG. 10 is a block diagram showing a schematic configuration of a voice recognition device which is a sixth embodiment of the present invention.

FIG. 11 is a flowchart showing a procedure for detecting a head loss.

FIG. 12 is a diagram for explaining the detection results of the environment estimation unit, the speech detection unit, and the speech defect detection unit shown in FIG. 10 and the response operation according to the detection results.

[Explanation of symbols]

1 Voice input section 2 Speech recognition section 2a recognition result 2b Cutout audio data 3 Situation detection section 4 Speech recognition result judgment unit 5 working parts 6 History analysis section 21 Speech recognition vocabulary storage 22 Rejected vocabulary storage 31 Environment Estimation Department 32 Speech detector 33 Speech defect detection unit

Claims (18)

[Claims] 1. A voice recognition unit for recognizing the input voice by collating the voice input unit for outputting voice data corresponding to the input voice with the standard pattern for the voice recognition vocabulary registered in advance. And a situation detecting means for checking whether or not at least one of a plurality of different factors causing a problem in the voice recognition exists in the voice data, and the situation detecting means detects any of the plurality of factors. If not, it is determined that the voice recognition is valid, and if any of the plurality of factors is detected, the voice recognition result determination means that determines that the voice recognition is not valid. And a plurality of response operations respectively corresponding to the plurality of factors are set in advance, and the voice recognition is not appropriate by the voice recognition result determination means. If it is constant, the speech recognition apparatus characterized by having a response unit for executing selectively a response operation to appropriate from among the plurality of response operations. 2. A rejection vocabulary storage unit in which a voice pattern relating to a rejection vocabulary for rejecting an input voice that is not an utterance based on the voice recognition vocabulary is registered, and the voice recognition unit includes the voice data and the rejection. The voice recognition device according to claim 1, wherein the voice recognition is performed by further collating with a voice pattern registered in the vocabulary storage means. 3. A history of factors detected by the situation detecting means is held, and when a factor is detected by the situation detecting means, the past occurrence state of the detected factor is checked from the history. Further comprising history analysis means for analyzing whether or not the examined occurrence state satisfies a predetermined condition, and the voice recognition result determination means further utilizes the analysis result in the history analysis means to validate the voice recognition. The speech recognition apparatus according to claim 1, wherein the response unit performs different response operations depending on whether the predetermined condition is satisfied or not. 4. The environment detecting means according to claim 1, wherein the situation detecting means includes an environment estimating means for detecting a cause of a defect due to a specific noise and an echo sound from the voice data. Voice recognition device. 5. The voice recognition apparatus according to claim 4, wherein the situation detection unit further includes an utterance detection unit that checks whether or not a human utterance is included in the voice data. 6. The utterance failure detecting means further comprises, when the utterance detecting means detects the utterance of the human being, checks whether or not a predetermined failure condition is satisfied for the detected utterance. The voice recognition device according to claim 5, comprising: 7. A first step of outputting voice data according to an input voice, and a second step of voice-recognizing the input voice by collating the voice data with a standard pattern relating to a voice recognition vocabulary registered in advance. And a third step of checking whether at least one of a plurality of different factors causing a problem in the voice recognition is present in the voice data, and the plurality of factors in the third step. If none of the above is detected, it is determined that the voice recognition is valid, and if any of the plurality of factors is detected, it is determined that the voice recognition is not valid. Four
And in the fourth step, when it is determined that the voice recognition is not appropriate, a corresponding response is selected from a plurality of preset response operations corresponding to the plurality of factors. A fifth step of selectively performing an operation, the voice recognition method. 8. The first step further comprises collating the voice data with a pre-registered voice pattern relating to a reject vocabulary for rejecting an input voice that is not a utterance based on the voice recognition vocabulary. The voice recognition method according to claim 7, further comprising the step of performing the voice recognition. 9. A step of holding a history of factors detected in the third step; and a step of detecting a factor in the third step, wherein the occurrence state of the detected factor in the past is recorded from the history. The method further includes a step of examining, and a step of analyzing whether or not the examined occurrence state satisfies a predetermined condition, and the fourth step further uses the result of the analysis to judge the validity of the voice recognition. 8. The voice recognition method according to claim 7, wherein the fifth step includes a step of performing different response operations depending on whether the predetermined condition is satisfied or not. 10. The method according to claim 7, wherein the third step includes a step of detecting a cause of a failure due to a specific noise and an echo sound from the voice data.
The voice recognition method according to any one of 1. 11. The voice recognition method according to claim 10, wherein the third step further includes a step of checking whether or not a human utterance is included in the voice data. 12. The third step further comprises, when the utterance of the human being is detected, checking whether or not a predetermined defect condition is satisfied for the detected utterance. 11. The voice recognition method according to item 11. 13. A first process of outputting voice data according to an input voice, and a second process of voice-recognizing the input voice by collating the voice data with a standard pattern relating to a voice recognition vocabulary registered in advance. Processing, a third processing for checking whether at least one of a plurality of different factors causing a problem in the voice recognition is present in the voice data, and the plurality of factors in the third processing. If none of the above is detected, it is determined that the voice recognition is valid, and if any of the plurality of factors is detected, it is determined that the voice recognition is not valid. 4 and a plurality of response actions corresponding respectively to the plurality of factors set in advance when the voice recognition is determined to be invalid in the fourth process. Program for executing a fifth process for performing responsive action by those selectively to the computer. 14. The first process further collates the voice data with a voice pattern, which is registered in advance, for a reject vocabulary for rejecting an input voice that is not a utterance based on the voice recognition vocabulary. 14. The program according to claim 13, including a process of performing the voice recognition. 15. A process of holding a history of factors detected by the third process; and a factor occurrence state in the past when the factor is detected by the third process, from the history. It further includes a process of examining and a process of analyzing whether or not the examined occurrence state satisfies a predetermined condition, and the fourth process further utilizes the result of the analysis to determine the validity of the voice recognition. 14. The program according to claim 13, wherein the fifth process includes a process of performing different response operations depending on whether the predetermined condition is satisfied or not. 16. The method according to claim 13, wherein the third process includes a process of detecting a cause of a problem caused by specific noise and echo sound from the voice data. program. 17. The program according to claim 16, wherein the third process further includes a process of checking whether or not a human utterance is included in the voice data. 18. The third process further includes a process of checking, when the utterance of the human being is detected, whether or not a predetermined defect condition is satisfied for the detected utterance. 17. The program according to 17.