JP2002215185A - Speech recognition system and voice recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition system which outputs a highly accurate recognized result, and to provide a voice recognition program. SOLUTION: The speech recognition system is provided with a microphone 102 for performing a voice input, an acoustic analysis part 103 for analyzing an inputted voice 101 inputted from the microphone 102, a sound/language processing part 105 which applies sound processing and language processing to the featured values series 104 of the analyzed sound to produce an answer candidate, and a sounding verification part 108 for calculating the probability of the answer candidate.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to a speech recognition system and a speech recognition program.

[0002]

2. Description of the Related Art In a speech recognition system, there have been proposed some methods for determining whether or not a recognition result is appropriate by using the certainty of an acoustic model or a language model.
For example, in the method of directly using the result of pattern matching with the model, it is determined whether or not the output score exceeds a predetermined threshold value, and only the one that exceeds the threshold value is determined as a result. Try to get recognition results. There is also a method in which a threshold value is set for the duration of the output score and information in the time direction is also used to obtain a more accurate recognition result.

[0003]

In the conventional method, the threshold value used as a criterion is a predetermined value, and the criterion does not take account of expansion and contraction in the time direction, so that the accuracy is limited. . That is, the value of the evaluation distance between the input voice and the standard pattern changes depending on the speaker and the environment, and the value and the duration are not constant even for the same speaker or environment. In this case, there is a problem that the determination accuracy is deteriorated. Further, even when the likelihood of a model to be referred to can be dynamically corresponded using an all-phoneme net or the like, the judgment scale uses a certain fixed threshold value and has the same problem.

An object of the present invention is to realize a speech recognition method and a speech recognition program capable of outputting a highly accurate recognition result.

[0005]

According to the speech recognition method of the present invention, an evaluation distance between a recognition result candidate and a model is calculated as a time series, and a time change characteristic is calculated from the evaluation distance series as a characteristic amount called a score change characteristic amount. Is extracted, and the score change feature amount is matched with the standard pattern of the score change feature amount in the correct answer and the error held in advance, and how high the likelihood that the correct answer model outputs to the error model is And outputting a recognition result candidate based on the calculated likelihood to obtain a highly accurate recognition result.

A speech recognition program according to the present invention stores a feature extraction process of analyzing input speech and extracting features, and features of the input speech extracted by the feature extraction process. An evaluation distance calculation process of calculating an evaluation distance by performing matching with a standard pattern of a voice feature amount; a candidate generation process of generating a correct answer candidate based on the evaluation distance calculated by the evaluation distance calculation process; An extraction process of extracting a score change feature amount representing a time change characteristic from an evaluation distance sequence for each frame determined at the time of calculation of the evaluation distance performed in the distance calculation process, and the score change feature amount extracted in the extraction process. ,
The accuracy calculation process of matching the score change feature amount of the correct answer example held in advance with the standard pattern and calculating the certainty of the correct candidate, and the correct answer based on the certainty of the correct candidate calculated in the certainty calculation process And causing the computer to execute a correct answer candidate output process of outputting candidates.

In the speech recognition method and the speech recognition program according to the present invention, the score change feature quantity serving as a criterion takes into account the time change of the evaluation distance, and furthermore, the correct distance model and the error model are created to create the evaluation distance. It is possible to use a criterion that is robust to the change in the absolute value of
It acts to output a highly accurate recognition result.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First, the principle of the speech recognition system of the present invention will be described. FIG. 4 is a characteristic diagram illustrating an example of a time change of the score change feature amount which is an evaluation distance. As shown in FIG. 4, the time change of the evaluation distance is 4 when the answer is correct.
In contrast to the case where the score is almost flat as indicated by 01, when an error is included, the score, which is the evaluation distance of the erroneous portion, may be locally reduced as indicated by reference numeral 402 due to an acoustic mismatch. Therefore, by modeling such a change in the evaluation distance with time, it can be considered that the correct answer candidate can be used to determine whether or not the answer candidate is really correct. In the modeling, a time series of scores is not directly used as a feature amount, but a feature amount extracted using a band-pass filter or the like is used to model an outline of the time series. When feature extraction is performed using a band-pass filter, the feature can be decomposed into frequency components. For example, it is easy to distinguish between a drop in a pulse-like score and a drop in a score over a certain section due to a mismatch. According to this determination method, correct / incorrect determination of a correct answer candidate that is robust not only in the absolute value of the score but also in expansion and contraction in the time direction can be performed. It should be noted that, in principle, even if only the correct answer model is modeled or an error model is used, it can be executed in principle.

Next, a speech recognition system according to a first embodiment of the present invention will be described in detail with reference to the drawings. Figure 1
FIG. 1 is a block diagram showing an overall configuration of a voice recognition system according to the first embodiment. This voice recognition method includes a microphone (input means) 102 for performing voice input, an acoustic analysis unit (feature extraction means) 103 for analyzing an input voice 101 input from the microphone 102, and an analyzed acoustic feature amount sequence 104. A sound / language processing unit (evaluation distance calculating means, candidate generating means, correct candidate output means) 105 for generating correct candidates by performing sound processing and language processing, and an utterance checking unit (probability calculating means) for calculating the probability of a correct candidate , Extraction means) 108. The sound / language processing unit 10
5 includes a hidden Markov model (H
MM), a statistical language model (N
Gram), and the utterance confirmation unit 108 is provided with a Gaussian mixture model (GMM) as a judgment model 109 in advance.

Next, the operation will be described with reference to FIG. An input voice 101 input from a microphone 102 is analyzed by a sound analysis unit 103 into a sound feature amount sequence 104. The sound / language processing unit 105 performs matching between the feature amount of the input voice and a standard pattern of the feature amount of the voice stored in advance, calculates an evaluation distance, and based on the calculated evaluation distance, a correct answer candidate. Generate That is, for the input acoustic feature amount sequence 104, a word sequence candidate that maximizes the product of the likelihood given by the acoustic model 106 and the likelihood given by the language model 107, and a plurality of candidates having the following likelihoods Generate candidates. Each candidate becomes a recognition result. At this time, the utterance confirmation unit 108 determines the likelihood of the correct answer candidate with respect to the likelihood time series 111 using the determination model 109, and among the determination results, Decision result 11 received at 108
Only the candidate corresponding to No. 2 is output as the recognition result 110.

Next, a detailed operation example of the utterance confirmation unit 108 will be described with reference to FIG. The utterance confirmation unit 108 receives the likelihood time series 111 of the correct answer candidates. The likelihood time series of the correct answer candidates is converted into a score change feature amount 202 by the band pass filter unit 201. That is, the sound / language processing unit 105 extracts a score change feature amount representing the time change characteristic from the evaluation distance sequence for each frame determined when calculating the evaluation distance. Then, the judgment model 203 compares the score change feature amount 202 with the judgment model 10.
9 is calculated.

The judgment model 109 includes a correct model, or a correct model and an error model. How high the likelihood that the correct model outputs is, or how much the likelihood that the correct model outputs with respect to the error model Whether the answer is high or not is determined. The determination result 112 is returned to the sound / language processing unit 105, and if accepted, the corresponding correct answer candidate is output. If rejected, the next candidate is processed.

FIG. 3 is a flow chart showing the determination processing operation in the utterance confirmation unit 108 shown in FIG. acoustic·
The likelihood time series 111 of the correct answer candidate is received from the language processing unit 105 (step S301), and the time change of the likelihood is extracted as a score change feature amount (step S302). A determination is made using a determination model prepared in advance (step S3
03), if the result of the determination is rejection (step S3)
04) Then, the processing proceeds to the next processing, and the processing from step S301 is performed on the next candidate. On the other hand, if the result of the determination in step S303 is accepted, the determination result 112 is output, and the process ends.

The above-described sound analysis unit 103, sound / language processing unit 105, and utterance confirmation unit 112 analyze the input speech and perform feature extraction, and the feature extraction process extracts the features. An evaluation distance calculation process of calculating the evaluation distance by performing matching between the feature amount of the input voice and a standard pattern of the feature amount of the voice held in advance, and an evaluation distance calculated by the evaluation distance calculation process. Candidate generation processing for generating a correct answer candidate based on the above, and extraction for extracting a score change feature amount representing a time change characteristic from an evaluation distance sequence for each frame determined at the time of evaluation distance calculation performed in the evaluation distance calculation processing Processing, the score change feature amount extracted in the extraction processing,
The accuracy calculation process of matching the score change feature amount of the correct answer example held in advance with the standard pattern and calculating the certainty of the correct candidate, and the correct answer based on the certainty of the correct candidate calculated in the certainty calculation process This can be realized as software as a speech recognition program for causing a computer to execute a correct answer candidate output process of outputting candidates.

As described above, according to the first embodiment, the likelihood is determined for the likelihood time series 111 of the correct answer candidates, and only those for which the determination result 112 is accepted are the recognition results 110. Therefore, there is an effect that a voice recognition method and a voice recognition program which can improve the accuracy of the output of the recognition result can be provided.

Next, a second embodiment of the present invention will be described. In the second embodiment, the absolute value of the output likelihood of the acoustic model is controlled by using the value normalized using the maximum value of the output likelihood of the acoustic model at that time as the score in FIG. It is possible to use a robust determination criterion that is difficult to be performed.

Next, a third embodiment of the present invention will be described. In the third embodiment, the output likelihood of the acoustic model and the score of the N-gram language model are used together as the score in FIG. Specifically, the output likelihood of the acoustic model and the value obtained by dividing and averaging the N-gram language model score for each frame are integrated for each frame to obtain a score. In the third embodiment, a more accurate judgment criterion can be used by using the language model score together.

[0018]

According to the present invention, as a result of the process of determining the likelihood of a correct answer candidate, a candidate that is not desirable for the application system or a candidate containing an undesired part can be rejected. There is an effect that becomes possible.

[Brief description of the drawings]

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

FIG. 2 is a block diagram illustrating a detailed configuration of an utterance confirmation unit of the voice recognition system according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing a determination processing operation in an utterance confirmation unit of the voice recognition system according to the first embodiment of the present invention.

FIG. 4 is a characteristic diagram illustrating an example of a time change of a score change feature amount which is an evaluation distance of the voice recognition method according to the first embodiment of the present invention.

[Explanation of symbols]

Reference numeral 102: microphone (input means), 103: sound analysis unit (feature extraction means), 104: sound feature amount sequence, 105: sound / language processing unit (evaluation distance calculation means,
Candidate generating means, correct answer candidate outputting means),... Utterance confirmation section (probability calculating means, extracting means).

Claims (6)

[Claims] 1. An input unit for inputting a voice, a feature extracting unit for analyzing a voice input by the input unit and extracting a feature, and a feature of the input voice extracted by the feature extracting unit An evaluation distance calculation unit that performs matching between the feature amount and a standard pattern of the feature amount of the voice held in advance, and calculates an evaluation distance; and generates a correct answer candidate based on the evaluation distance calculated by the evaluation distance calculation unit. Candidate generating means, and extracting means for extracting, from the evaluation distance sequence for each frame determined at the time of calculating the evaluation distance performed by the evaluation distance calculating means, a score change feature representing the time change characteristic thereof; Matching of the extracted score change feature quantity with the standard pattern of the score change feature quantity of the correct answer example held in advance is performed, and the correct answer candidate is confirmed. And probability calculating means for calculating and of speech recognition system in which the correct candidate output means, and comprising the outputting the correct candidates based on likelihood of correct candidates the probability calculating means is calculated. 2. The speech recognition method according to claim 1, wherein not only the correct answer example but also the score change feature amount of the error example are used together as the standard pattern for matching. 3. The method according to claim 1, wherein the temporal change in the evaluation distance over the entire utterance is analyzed using a band-pass filter group, and the characteristic amount is used as the characteristic amount of the temporal change in the evaluation distance. The described speech recognition method. 4. The speech recognition method according to claim 1, wherein a value normalized using a maximum value of the output likelihood of the acoustic model at that time is used as the evaluation distance. 5. The speech recognition method according to claim 1, wherein said evaluation distance calculating means uses an output likelihood of an acoustic model and a score of an N-gram language model as an evaluation distance. 6. A feature extraction process for analyzing an input voice to extract features, and a standard of a feature value of the input voice extracted by the feature extraction process and a feature value of a voice stored in advance. An evaluation distance calculation process of calculating an evaluation distance by performing matching with a pattern; a candidate generation process of generating a correct answer candidate based on the evaluation distance calculated by the evaluation distance calculation process; and an evaluation performed by the evaluation distance calculation process. An extraction process of extracting a score change feature representing a time change characteristic from an evaluation distance sequence for each frame determined at the time of calculating the distance, and the score change feature extracted in the extraction process are stored in advance. The accuracy calculation process of matching the score change feature amount of the correct answer example with the standard pattern and calculating the likelihood of the correct answer candidate, and the accuracy calculation process Speech recognition program for executing a correct candidate output process of outputting the correct candidates based on likelihood of correct candidates out, to a computer.