JP4583772B2 - Speech recognition system, speech recognition method, and speech recognition program - Google Patents

Description

  The present invention relates to a voice recognition system, a voice recognition method, and a voice recognition program for recognizing the voice of an unspecified speaker.

  A voice recognition system that recognizes the voice of an unspecified speaker using an unspecified speaker acoustic model does not require voice registration by the user, but an acoustic model specialized for a specific speaker or speaker group is used. The recognition performance generally deteriorates compared to the case of using. Here, the acoustic model is a model that gives an acoustic certainty to the input voice.

  In addition, there is a speech recognition system that recognizes a speaker selected from a plurality of speakers after determining a plurality of speakers using a GMM (Gaussian Mixture Model) or the like before the speech recognition. However, in this speech recognition system, the amount of processing increases as much as the preprocessing is performed, and the input speech must be processed once during the preprocessing, so that the response is delayed.

  In view of this, there is a speech recognition system that recognizes a plurality of speakers' acoustic models in parallel, and finally selects one having a good speech recognition score, that is, discriminates a plurality of speakers. However, this speech recognition system requires a calculation amount corresponding to the number of speakers.

  On the other hand, Patent Document 1 discloses an example of a speech recognition system that reduces the amount of calculation. This speech recognition system is a system that automatically selects the best speaker by commonly performing a process called pruning described later for a plurality of hypotheses corresponding to a plurality of speakers, as shown in FIG. , Microphone 1, feature extraction unit 2, buffer memory 3, phoneme matching unit 4, LR parser 5, speaker mixed hidden Markov network memory 11, speaker model memory 12, LR table memory 13, And a context-free grammar database memory 20.

  The conventional speech recognition system having such a configuration operates as follows.

  First, a speaker voice is received by the microphone 1, and a feature amount that is a parameter of the speaker voice is extracted by the feature extraction unit 2. The extracted feature amount is input to the phoneme matching unit 4 via the buffer memory 3.

  In the phoneme matching unit 4, phoneme matching of the input feature quantity is performed in response to a phoneme matching request from the LR parser 5. At this time, the phoneme context information including information on the phoneme collation section and information on the phonemes to be collated and the phonemes before and after the phoneme is passed from the LR parser 5. In the phoneme matching unit 4, a state on a hidden Markov network (hereinafter referred to as an HM network) that can accept such a phoneme context based on the received phoneme context information is defined as a restriction between the preceding state list and the succeeding state list. A model is selected by concatenating within. Then, the likelihood for the data in the phoneme matching section is calculated using the selected model, and the likelihood value is returned to the LR parser 5 as a phoneme matching score.

  The LR parser 5 refers to, for example, the speaker model memory 12 including the phoneme duration model and the LR table memory 13, and performs hypothesis calculation processing for the input phoneme matching score from left to right without backtracking. I do. The LR table memory 13 stores an LR table in which a predetermined context free grammar (CFG) in the context free grammar database memory 20 is automatically converted. The LR parser 5 predicts the next phoneme from the LR table in the LR table memory 13 and outputs the phoneme prediction data to the phoneme matching unit 4.

  In response to this, the phoneme collation unit 4 refers to the information in the hidden Markov network memory 11 corresponding to the input phoneme prediction data and collates it, and returns the likelihood to the LR parser 5 as a speech recognition score.

  The LR parser 5 performs continuous speech recognition by sequentially connecting phonemes based on the input speech recognition score. Note that when a plurality of phonemes are predicted by the phoneme matching unit 4, high-speed processing is realized by checking the existence of all of them and performing pruning to leave a partial tree with a high likelihood of partial speech recognition. .

After that, when the speaker voice received by the microphone 1 is processed to the end, the LR parser 5 outputs the one with the maximum overall likelihood or a plurality of predetermined higher ranks as voice recognition result data.
Japanese Patent No. 2905673

  However, as described above, when speech recognition is performed while automatically identifying multiple speakers, in a speech recognition system that prunes hypotheses of multiple speakers in common, processing up to hypothesis calculation processing is performed for all speakers. Since it must be performed on the model, there is a problem that it is not efficient in terms of processing amount.

  An object of the present invention is to provide a speech recognition system, a speech recognition method, and a speech recognition program capable of performing speech recognition of an input speech while discriminating and selecting an acoustic model of the input speech with a small processing amount. is there.

The speech recognition system according to the present invention includes an acoustic likelihood calculating means for calculating acoustic likelihoods with a plurality of acoustic models for the feature quantity of the input speech, and a predetermined interval at the head of the input speech by the acoustic likelihood calculating means. Model discrimination means for discriminating and selecting the acoustic model of the input speech based on the acoustic likelihood calculated by the acoustic likelihood calculation means within the predetermined section at When the acoustic likelihood calculation for the predetermined interval of the input speech is completed by the likelihood calculating means, hypothesis processing is performed using the acoustic likelihood calculated for the acoustic model selected by the model discriminating means within the predetermined interval. And a hypothesis processing means for outputting a speech recognition result based on the hypothesis processing result after the hypothesis processing for all sections of the input speech is completed. Here, the acoustic likelihood calculating means calculates the acoustic likelihood with all of the plurality of acoustic models for the feature amount of the input speech in a predetermined section at the beginning of the input speech, and determines the predetermined values after the top of the input speech. In the section, the acoustic likelihood with only the acoustic model selected by the model discriminating means is calculated for the feature quantity of the input speech.

  With such a configuration, the acoustic likelihood calculation process is performed a plurality of times for the number of the plurality of acoustic models only in the first predetermined section of the input speech, and the acoustic likelihood calculation process has already been selected in the predetermined section after the start. It is performed only for the number of acoustic models that are being used, that is, only once. Therefore, it is possible to reduce the processing amount when performing speech recognition of the input speech while discriminating and selecting the acoustic model of the input speech.

  Here, the model discrimination target may be a speaker, and the speaker discrimination target in that case may be gender, age, or dialect. Alternatively, the model discrimination target may be a noise environment, transmission characteristics, language, or a combination thereof.

  According to the present invention, the acoustic likelihood calculation process is performed a plurality of times for the number of the plurality of acoustic models only in the first predetermined section of the input speech, and the acoustic likelihood calculation process is already selected in the predetermined section after the start. It is performed only for the number of acoustic models that are used, that is, only once, and at the time of hypothesis processing, the acoustic likelihood calculated during the acoustic likelihood calculation processing of the previous stage is shared, and no new acoustic likelihood is calculated. Therefore, it is possible to reduce the processing amount when performing speech recognition of the input speech while discriminating and selecting the acoustic model of the input speech.

  Next, the best mode for carrying out the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing the configuration of a speech recognition system according to an embodiment of the present invention. The speech recognition system includes an acoustic model storage unit 105 that stores a plurality of acoustic models, a language model storage unit 107 that stores a language model, a microphone 102 that performs speech input of the input speech 101, and parameters of the input speech 101. A feature extraction unit 103 that extracts a feature amount, an acoustic likelihood calculation unit 104 that calculates acoustic likelihoods for a plurality of acoustic models for the extracted feature amount, and a predetermined interval at the beginning of the input speech 101 When the acoustic likelihood calculation for the input speech 101 is completed, the model discrimination unit 108 for discriminating and selecting the acoustic model of the input speech 101 based on the acoustic likelihood calculated in the predetermined interval, and the predetermined speech segment of the input speech 101 Each time the acoustic likelihood calculation is completed, hypothesis processing is performed using the acoustic likelihood calculated for the acoustic model selected in the predetermined section, and all of the input speech 101 is calculated. After hypothesis processing between is completed, and a hypothesis processor 306 which outputs a speech recognition result 109 hypothetical processing result. Here, the acoustic likelihood calculation unit 104 calculates the acoustic likelihood with all of the plurality of acoustic models for the feature amount of the input speech 101 in the first predetermined section of the input speech 101. In a predetermined section, the acoustic likelihood with only the acoustic model selected by the model discrimination unit 108 is calculated for the feature amount of the input speech 101.

  Next, an example of the operation of the speech recognition system according to the present embodiment will be described using the flowchart of FIG.

  First, the input voice 101 is input from the microphone 102 in time synchronization, for example, every 10 milliseconds (step 201). The sample amount for 10 milliseconds is defined as one frame.

  Next, the feature extraction unit 103 extracts the feature value of one frame of the input sound 101 (step 202).

  Next, in the acoustic likelihood calculation unit 104, the acoustic likelihood with the plurality of acoustic models stored in the acoustic model storage unit 105 is calculated with respect to the feature amount for one frame extracted by the feature extraction unit 103. (Step 203). At this time, in the acoustic likelihood calculation unit 104, the acoustic likelihood with all of the plurality of acoustic models is calculated in the predetermined section at the head of the input speech 101, but in the subsequent sections, the already selected acoustics are calculated. Calculate acoustic likelihood with model only. The acoustic likelihood calculated here is stored in the acoustic likelihood calculation unit 104.

  Here, when the acoustic likelihood is accumulated for a predetermined section, for example, 60 frames, if this section is the head of the input speech 101 (Yes in Step 204), the model discrimination unit 108 calculates the acoustic likelihood. The acoustic model of the input speech 101 is discriminated and selected based on (step 205), and the acoustic likelihood calculated by the acoustic likelihood calculator 104 for the selected acoustic model is the hypothesis processor 106. Used for hypothesis processing in On the other hand, if it is other than the head of the input speech 101 (No in step 204), the model discrimination by the model discrimination unit 108 is not performed, and the acoustic likelihood calculation unit for the acoustic model already selected by the model discrimination unit 108 The acoustic likelihood calculated at 104 is used for hypothesis processing in the hypothesis processing unit 106.

  Next, in the hypothesis processing unit 106, the acoustic likelihood calculated in the predetermined section by the acoustic likelihood calculation unit 104 for the acoustic model selected by the model determination unit 108 and the language model storage unit 107. Hypothesis processing is performed frame by frame using the stored language model (step 206).

  When the process up to the hypothesis processing for a predetermined section of the input speech 101 is completed, if the end of the section is not the end of all sections of the input speech 101 (No in Step 207), the process returns to Step 201, Accept next frame input. On the other hand, if it is the end of all sections of the input speech 101 (Yes in Step 207), the hypothesis processing unit 106 outputs the most probable word sequence as the speech recognition result 109 from the word sequences obtained by the hypothesis processing. (Step 208), the process ends.

  As described above, in the present invention, in the first predetermined section of the input speech 101, the acoustic likelihood calculation processing is performed by the acoustic likelihood calculation unit 104 using all the plurality of acoustic models, and the acoustic likelihood of the first predetermined section is calculated. When the degree calculation process is completed, the model determination unit 108 selects an acoustic model with a good acoustic likelihood calculation result. In a predetermined section after the head, the acoustic likelihood calculation process is performed by the acoustic likelihood calculation unit 104 using only the acoustic model selected by the model determination unit 108.

  Thereby, the acoustic likelihood calculation process is performed a plurality of times only for the number of the plurality of acoustic models only in the head section of the input speech 101, and in the section after the head, the acoustic likelihood calculation process is already selected. Therefore, it is possible to reduce the amount of processing when the acoustic model of the input speech 101 is selected from a plurality of acoustic models and speech recognition is performed.

  Next, the speech recognition system of the present invention will be described using specific examples.

(First embodiment)
FIG. 3 is a block diagram showing the configuration of the voice recognition system according to the first embodiment of the present invention. This speech recognition system uses the speech of an unspecified speaker as a specific example of the input speech 101 shown in FIG. 1, thereby making the model discrimination target a speaker, and a male speaker model as a specific example of an acoustic model. And the female speaker model, the target of speaker discrimination is gender. Accordingly, the model discriminating unit 108 shown in FIG. 1 is replaced with a speaker discriminating unit 308 that discriminates a speaker model, and the acoustic model storage unit 105 is a male speaker model storage that stores a male speaker model. It replaces with the part 305A and the female speaker model storage part 305B which stores a female speaker model. In FIG. 3, the microphone 302, the feature extraction unit 303, the acoustic likelihood calculation unit 304, the hypothesis processing unit 306, and the language model storage unit 307 are respectively the microphone 102, the feature extraction unit 103, and the acoustic likelihood shown in FIG. This corresponds to the degree calculation unit 104, hypothesis processing unit 106, and language model storage unit 107.

  Next, an example of the operation of the voice recognition system according to this embodiment will be described.

  First, the input sound 301 is input from the microphone 302 in time synchronization, for example, every 10 milliseconds. The sample amount for 10 milliseconds is defined as one frame.

  Next, the feature extraction unit 303 extracts the feature amount of one frame of the input sound 301.

  Next, in the acoustic likelihood calculation unit 304, the acoustic likelihood with the male speaker model stored in the male speaker model storage unit 305A with respect to the feature amount for one frame extracted by the feature extraction unit 303, The acoustic likelihood with the female speaker model stored in the female speaker model storage unit 305B is calculated. The acoustic likelihood calculated here is accumulated in the acoustic likelihood calculation unit 304.

  Here, when the acoustic likelihood for the first predetermined section of the input speech 301, for example, the first 60 frames, is accumulated, the speaker discrimination unit 308 determines the male speaker model and female speaker model for the section. Compare the average values of each acoustic likelihood, for example, if the average acoustic likelihood of the male speaker model is more likely than the average acoustic likelihood of the female speaker model, select the male speaker model . In the following description, it is assumed that the male speaker model is selected.

  Next, in the hypothesis processing unit 306, the acoustic likelihood calculated in the predetermined section by the acoustic likelihood calculation unit 304 for the male speaker model selected by the speaker determination unit 308, and the language model storage Hypothesis processing is performed frame by frame using the language model stored in the unit 307, and a plurality of word string hypotheses are obtained in a probable order.

  When the processing up to the hypothesis processing for the first predetermined section of the input speech 301 is completed, the acoustic likelihood calculation process and the hypothesis process are respectively performed using only the already selected male speaker model for the subsequent sections. When the degree calculation unit 304 and the hypothesis processing unit 306 complete the processing up to the end of all sections of the input speech 301, the hypothesis processing unit 306 outputs the most likely word string as the speech recognition result 309.

  As described above, in this embodiment, it is possible to use a speaker model specialized for men or women without specifying the sex of the speaker in advance, and even if compared with the case where the gender is specified in advance. Less increase in throughput. That is, a plurality of acoustic likelihood calculation processes (in this embodiment, twice) are performed only in the first predetermined section.

  In this embodiment, the speaker discrimination target is gender, and the speaker is divided into two clusters of male and female, but the speaker is divided into a plurality of clusters according to some other criteria, It is also possible to easily determine the speaker model. For example, if the target of speaker discrimination is classified by age and child speech, adult speech and elderly speech are discriminated, or the speaker discrimination target is distinguished by dialect, distinguishing between American English and Japanese English If you want to.

  In this embodiment, the model discrimination target is a speaker, but the model discrimination target is not limited to a speaker. For example, a model discrimination target is a noise environment, and speech recognition is performed while determining the noise environment using models corresponding to a plurality of noise environments having different S / N ratios. There may be a case where voice recognition is performed while discriminating transmission characteristics using models corresponding to different transmission characteristics such as voice and fixed telephone voice. In addition, it is conceivable that model recognition is performed on a language and, for example, three types of models of Japanese, English, and Chinese are used, and speech recognition is performed while performing language determination. At this time, a method of narrowing the language to one of the predetermined intervals at the beginning may be adopted, or two probable ones are first selected and finally limited to one in the next predetermined interval. You may adopt the method of. Further, a model discrimination target may be a combination of the above-described noise environment, transmission characteristics, and language.

  Furthermore, in this embodiment, as a method of comparing the acoustic likelihood for speaker discrimination, a method using the average value of the frame unit acoustic likelihood for the section is adopted. However, the present invention is not limited to this method. A method may be adopted in which superiority or inferiority of acoustic likelihood is compared for each and a majority decision is made for that section.

(Second embodiment)
FIG. 4 is a block diagram showing the configuration of a speech recognition system according to the second embodiment of the present invention. As in the first embodiment, this speech recognition system uses the speech of an unspecified speaker as a specific example of the input speech 101 shown in FIG. By using a male speaker model and a female speaker model as specific examples of the model, the target of speaker discrimination is set to gender. Accordingly, the model discriminating unit 108 shown in FIG. 1 is replaced with a speaker discriminating unit 408 that discriminates a speaker model, and the acoustic model storage unit 105 is a male speaker model storage that stores a male speaker model. It replaces with the part 405A and the female speaker model storage part 405B which stores a female speaker model. In FIG. 4, the microphone 402, the feature extraction unit 403, the acoustic likelihood calculation unit 404, the hypothesis processing unit 406, and the language model storage unit 407 are the microphone 102, the feature extraction unit 103, and the acoustic likelihood shown in FIG. This corresponds to the degree calculation unit 104, hypothesis processing unit 106, and language model storage unit 107. Further, the speech recognition system is further provided with a threshold determination unit 410 that determines whether the average acoustic likelihood of the model to be selected by the speaker determination unit 408 exceeds a predetermined threshold.

  Next, an example of the operation of the voice recognition system according to this embodiment will be described.

  First, the input sound 401 is input from the microphone 402 in time synchronization, for example, every 10 milliseconds. The sample amount for 10 milliseconds is defined as one frame.

  Next, the feature extraction unit 403 extracts the feature amount of the sound for one frame of the input sound 401.

  Next, in the acoustic likelihood calculation unit 404, the acoustic likelihood with the male speaker model stored in the male speaker model storage unit 405A with respect to the feature amount of one frame extracted by the feature extraction unit 403, The acoustic likelihood with the female speaker model stored in the female speaker model storage unit 405B is calculated. The acoustic likelihood calculated here is accumulated in the acoustic likelihood calculation unit 404.

  Here, when the acoustic likelihood for the first predetermined section of the input speech 401, for example, the first 30 frames, is accumulated, the speaker discrimination unit 408 stores the male speaker model and the female speaker model for the section. Compare the average values of the respective acoustic likelihoods.For example, if the average acoustic likelihood of the male speaker model is more likely than the average acoustic likelihood of the female speaker model, the average acoustic likelihood of the male speaker model The likelihood is passed to the threshold determination unit 410. In the following description, it is assumed that the average acoustic likelihood of the male speaker model is passed to the threshold determination unit 410.

  Next, the threshold determination unit 410 determines whether the average acoustic likelihood of the male speaker model passed from the speaker determination unit 408 exceeds a predetermined threshold. Is returned to the speaker discrimination unit 408. In the following description, it is assumed that the determination result of acceptance is returned to the speaker determination unit 408.

  Next, since the determination result of acceptance is given from the threshold value determination unit 410 in the speaker determination unit 408, a male speaker model is selected as the model determination result.

  Next, in the hypothesis processing unit 406, the acoustic likelihood calculated in the predetermined section by the acoustic likelihood calculation unit 404 for the male speaker model selected by the speaker determination unit 408, and the language model storage Hypothesis processing is performed frame by frame using the language model stored in the unit 407, and a plurality of word string hypotheses are obtained in a probable order.

  When the process up to the hypothesis process for the first predetermined section of the input speech 401 is completed, the acoustic likelihood calculation process and the hypothesis process are respectively performed using only the already selected male speaker model for the subsequent sections. When the degree calculation unit 404 and the hypothesis processing unit 406 complete the processing up to the end of the entire section of the input speech 401, the hypothesis processing unit 406 outputs the most likely word string as the speech recognition result 409.

  As described above, in the present embodiment, by performing the threshold determination, it is possible to make the head section in which the speaker discrimination is performed in the section of the input speech 401 variable, and as a result, more accurate speech. Person identification can be performed.

  In the present invention, the processing in the voice recognition system may be executed by a voice recognition program for realizing the function in addition to the above-described dedicated hardware. In this case, the voice recognition system is configured as one computer, the voice recognition program is recorded on a recording medium readable by the computer, and the voice recognition program recorded on the recording medium is read and executed by the computer. . The computer-readable recording medium refers to a transfer medium such as a floppy disk, a magneto-optical disk, a DVD, and a CD, as well as an HDD built in the computer.

  The present invention is applicable to uses such as a voice input interface for a computer.

It is a block diagram which shows the structure of the speech recognition system by one Embodiment of this invention. It is a flowchart explaining an example of operation | movement of the speech recognition system shown in FIG. It is a block diagram which shows the structure of the speech recognition system by 1st Example of this invention. It is a block diagram which shows the structure of the speech recognition system by 2nd Example of this invention. It is a block diagram which shows an example of a structure of the conventional speech recognition system.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Input speech 102 Microphone 103 Feature extraction part 104 Acoustic likelihood calculation part 105 Acoustic model storage part 106 Hypothesis processing part 107 Language model storage part 108 Model discrimination | determination part 109 Speech recognition result 201-208 Step 301 Input voice 302 Microphone 303 Feature extraction part 304 Acoustic likelihood calculation unit 305A Male speaker model storage unit 305B Female speaker model storage unit 306 Hypothesis processing unit 307 Language model storage unit 308 Speaker discrimination unit 309 Speech recognition result 401 Input speech 402 Microphone 403 Feature extraction unit 404 Acoustic likelihood Degree calculation unit 405A Male speaker model storage unit 405B Female speaker model storage unit 406 Hypothesis processing unit 407 Language model storage unit 408 Speaker discrimination unit 409 Speech recognition result 410 Threshold determination unit

Claims (12)

In a speech recognition system that recognizes the input speech while distinguishing and selecting the acoustic model of the input speech from a plurality of acoustic models,
Acoustic likelihood calculating means for calculating the acoustic likelihood with the plurality of acoustic models in time synchronization with respect to the feature amount of the input speech;
Based on the acoustic likelihood calculated by the acoustic likelihood calculation means within the predetermined interval at the time when the acoustic likelihood calculation for the predetermined interval at the beginning of the input speech is completed by the acoustic likelihood calculation means. Model discrimination means for discriminating and selecting an acoustic model of the input speech;
Wherein the acoustic likelihood calculating means with the input speech in a predetermined time section acoustic likelihood calculation is finished, the acoustic likelihood calculated for the acoustic model selected by the model discriminating unit to the predetermined interval within A hypothesis processing means for performing a hypothesis processing in time synchronization using the, and after completing the hypothesis processing of all sections of the input speech, and outputting a speech recognition result based on the hypothesis processing result,
The acoustic likelihood calculating means calculates the acoustic likelihood with all of the plurality of acoustic models for the feature amount of the input speech in a predetermined section at the beginning of the input speech in a time-synchronized manner. In a predetermined section after the head of the speech recognition system, an acoustic likelihood with only the acoustic model selected by the model discrimination means is calculated in time synchronization with respect to the feature quantity of the input speech. The speech recognition system according to claim 1 , wherein a model discrimination target by the model discrimination means is a speaker. The speech recognition system according to claim 2 , wherein a target of speaker discrimination by the model discrimination means is gender, age, or dialect. The speech recognition system according to claim 1 , wherein the model discrimination target by the model discrimination unit is a noise environment, transmission characteristics, language, or a combination thereof. In a speech recognition method by a speech recognition system that performs speech recognition of the input speech while discriminating and selecting an acoustic model of the input speech from a plurality of acoustic models,
A first step in which an acoustic likelihood calculating means calculates the acoustic likelihood with all of the plurality of acoustic models for the feature amount of the input speech in a predetermined interval at the beginning of the input speech in a time-synchronized manner; ,
When the model discrimination means finishes calculating the acoustic likelihood for the predetermined interval at the beginning of the input speech, the input speech is calculated based on the acoustic likelihood calculated in the first step within the predetermined interval. A second step of determining and selecting an acoustic model of
When the hypothesis processing means completes the acoustic likelihood calculation for the predetermined interval at the beginning of the input speech, the hypothesis processing means performs the first step on the acoustic model selected in the second step within the predetermined interval. A third step of performing hypothesis processing in time synchronization using the calculated acoustic likelihood;
The acoustic likelihood calculation means calculates the acoustic likelihood with only the acoustic model selected in the second step with respect to the feature quantity of the input voice in a predetermined section after the head of the input voice in time synchronization. A fourth step to:
When the hypothesis processing means finishes calculating the acoustic likelihood for a predetermined section after the head of the input speech, the hypothesis processing means performs the fourth step on the acoustic model selected in the second step within the predetermined section. A fifth step of performing a hypothesis process in time synchronization using the calculated acoustic likelihood and outputting a speech recognition result based on the hypothesis process result after the hypothesis process of all sections of the input speech is completed. A voice recognition method characterized by the above. The speech recognition method according to claim 5 , wherein in the second step, a model discrimination target is a speaker. The speech recognition method according to claim 6 , wherein in the second step, a speaker discrimination target is gender, age, or dialect. The speech recognition method according to claim 5 , wherein in the second step, a model discrimination target is a noise environment, a transmission characteristic, a language, or a combination thereof. In a speech recognition program to be executed by a computer that performs speech recognition of the input speech while determining and selecting an acoustic model of the input speech from a plurality of acoustic models,
A first step of calculating, with time synchronization, acoustic likelihoods with all of the plurality of acoustic models for a feature amount of the input speech in a predetermined section at the beginning of the input speech;
When the acoustic likelihood calculation for the first predetermined section of the input speech is completed, the acoustic model of the input speech is determined based on the acoustic likelihood calculated in the first step within the predetermined section. A second step of making a selection;
The acoustic likelihood calculated in the first step with respect to the acoustic model selected in the second step within the predetermined interval when the calculation of the acoustic likelihood for the predetermined interval at the beginning of the input speech is completed. A third step of performing hypothesis processing in time synchronization using degrees;
A fourth step of calculating, in time synchronization, an acoustic likelihood of only the acoustic model selected in the second step with respect to a feature quantity of the input speech in a predetermined section after the head of the input speech;
When the acoustic likelihood calculation for a predetermined section after the head of the input speech is completed, the acoustic likelihood calculated in the fourth step with respect to the acoustic model selected in the second step within the predetermined section. performs hypothesis treated with time synchronization with, after the hypothesis processing of all segments of the input speech is finished, characterized in that to execute a fifth step of outputting the speech recognition result based on the hypothesis processing result to the computer Voice recognition program. The speech recognition program according to claim 9 , wherein in the second step, a model discrimination target is a speaker. The speech recognition program according to claim 10 , wherein in the second step, a speaker discrimination target is gender, age, or dialect. The speech recognition program according to claim 9 , wherein in the second step, a model discrimination target is a noise environment, a transmission characteristic, a language, or a combination thereof.

Images