JP2001296884A - Device and method for voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device and a method for voice recognition to which the condition of uttering of a system user and external factors such as environmental condition are introduced as new knowledge sources. SOLUTION: In the voice recognition device which analyzes inputted voice into acoustic parameters and recognizes the voice by comparing the parameters with comparison-object pattern candidates, the likelihood of recognition is weighted employing the values that are set based on the time of the uttering and external factors to conduct converging of word recognition. Thus, a more likely result or a higher recognition rate is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention analyzes an input speech and compares it with a plurality of comparison patterns input in advance.
The present invention relates to a speech recognition device and method for recognizing speech using a knowledge source such as linguistic information.

[0002]

2. Description of the Related Art In recent years, one of man-machine interfaces has been developed.
As a method, a speech dialogue interface using speech recognition has attracted attention. As a conventional technique, in addition to an acoustic evaluation value, a linguistic evaluation value generated by a statistical language model (finite automaton, stochastic context-free grammar) represented by an N-gram is weighted. In addition, there is a method of restricting a search vocabulary to speed up recognition and improve a recognition rate (for example, JP-A-09-134192 and JP-A-09-274498). Further, as a conventional technique, a method of changing a language model in a dialogue state (Japanese Unexamined Patent Application Publication No.
No. 07-104786) has also been proposed.

[0003]

In the conventional speech recognition apparatus, the probability of transition from some words uttered immediately before to the next word is statistically given, so that it depends only on the grammar of the language. , Does not take into account the state of the speaker. Further, since the learning uses static data, the state of the speaker may be temporally and psychologically different from the time of learning when actually used as a system. Therefore, it was not possible to pursue a higher recognition rate using only the conventional language model. Speech recognition devices are attracting attention as future man-machine interfaces. However, there is a problem that a higher recognition rate cannot be obtained with a knowledge source using only a language model as a calculation constraint condition at the time of recognition. Another problem is that a new knowledge source including the state and environment of the speaker is required. An object of the present invention is to solve the above problem. That is, the object of the present invention is:
It is an object of the present invention to provide a speech recognition apparatus and method in which external factors other than the utterance of the system user, such as the state and environment, are introduced as new knowledge sources.

[0004]

In order to solve the above-mentioned problems, the invention according to the first aspect analyzes input speech into acoustic parameters, and stores a plurality of comparison target pattern candidates stored in advance. In comparison with, the voice recognition device that recognizes voice focuses on the temporal bias of the uttered word,
Means for obtaining a value set based on the utterance time;
Means for weighting the likelihood after speech recognition obtained in advance using the obtained value and narrowing down likely words. According to a second aspect of the present invention, in the speech recognition apparatus according to the first aspect, the likelihood after speech recognition is determined in advance by using an external factor and paying attention to an external factor using a value set for the external factor. Is performed. According to a third aspect of the present invention, there is provided a speech recognition apparatus for analyzing an input speech into acoustic parameters, comparing the analysis result with a plurality of comparison target pattern candidates stored in advance, and recognizing the speech. Focusing on the temporal bias of the uttered word, obtaining a value set based on the uttered time, and weighting the likelihood after speech recognition obtained in advance using the obtained value. , To narrow down likely words. According to a fourth aspect of the present invention, in the speech recognition method according to the third aspect, the likelihood after speech recognition is determined in advance by using an external factor and using a value set for the external factor. Is performed.

[0005]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a voice recognition device according to one embodiment of the present invention. As shown in FIG. 1, the speech recognition device includes a microphone 1, a speech feature extraction unit 2, a speech recognition unit 3, a phoneme HMM 4, a word dictionary 5, a word hypothesis narrowing unit 6, a statistical language model 7, And a speaker state analysis unit 8 and an utterance knowledge base 9. The audio feature extraction unit 2 is connected to the microphone 1. The speech recognition unit 3 is connected to the speech feature extraction unit 2. The phoneme HMM 4 and the word dictionary 5 are connected to the speech recognition unit 3. Word hypothesis refiner 6
Are connected to the voice recognition unit 3. The statistical language model 7 is connected to the word hypothesis narrowing unit 6. The speaker state analysis unit 8 is connected to the word hypothesis narrowing unit 6. The utterance knowledge base 9 is connected to the utterer state analysis unit 8. The microphone 1 is a voice input unit for inputting voice (hereinafter, referred to as a microphone). The voice feature extraction unit 2 analyzes the input voice into a feature parameter. In the speech recognition unit 3, the phoneme HMM4 (HMM: Hidden Markov Mode in this embodiment), which is model data of acoustic parameters
l is used) and the word dictionary 5 which is model data of grammatical parameters is used for comparison and recognition. The utterance knowledge base 9 stores words to be used, times at which the words are frequently used, and the like. In the utterance knowledge base 9, in addition to the information on the temporal factors, information on other external factors is similarly stored for each factor. The speaker state analysis unit 8 analyzes the state from the environment at the time of the speaker's utterance, such as the current time, based on the utterance knowledge base 9 and indicates the possibility of being used in the current recognition target vocabulary. This is a part for calculating an evaluation value. The word hypothesis narrowing unit 6 compares the result calculated by the speaker state analysis unit 8 with the statistical language model 7 and the speech recognition unit 3 to narrow down words.

[0006] Next, the process from the voice input to the recognition result output of the apparatus will be described. For example, when saying “What is it?” To the microphone 1, the MFCC (mel frequency cepstral coefficient) is calculated by the voice feature extraction unit 2 and sent to the voice recognition unit 3. Based on the MFCC, the speech recognition unit 3 detects a word hypothesis using the phoneme HMM 4 and the word dictionary 5, and calculates the likelihood. For example, the recognition likelihood from the speech recognition unit 3 is calculated as the sum of the forward path of the HMM multiplied by the output probability and the transition probability. The calculated result is output as shown in FIG.
In the present embodiment, the utterance “chosho” is acoustically difficult to distinguish from “chosho”. Also, grammatically, "what is breakfast" and "what is lunch" are both composed of S + V (subject and predicate), and these two sentences are grammatically identical. Since there is no difference between language models and they cannot be distinguished, when the hypothesis is narrowed down by the word hypothesis narrowing unit 6, even if the statistical language model 7 is used, the word cannot be completely narrowed down. Therefore, the parameters from the speaker state analysis unit 8 are used. Speaker state analysis unit 8
The function of is shown below. The speaker state analyzing unit 8 calculates a parameter for weighting the likelihood calculated by the speech recognition unit 3 by associating the utterance knowledge base 9 with time and other external factors. The contents of the utterance knowledge base 9 in this example are shown in FIG.
In the case of a format like Utterance knowledge base 9 of FIG.
Is a table in which each element of the word used, the start time A and the end time B of the section in which the word is frequently used, and the evaluation value (weight) at that time are stored.
The utterer state analysis unit 8 captures the current time, and if any of the word hypotheses output from the speech recognition unit 3 is registered in the utterance knowledge base 9, the evaluation value of the section corresponding to the current time Is transferred to the word hypothesis narrowing unit 6. For those not registered, the lowest value of the utterance knowledge base 9 is set, and the value is transferred to the word hypothesis narrowing unit 6. The invention according to claims 1 and 3 of the present invention resides in introducing the utterance probability of a word using time as a variable as an evaluation value other than the language model. In conversation, the utterance probabilities of words used are biased with respect to time (date and time). For example, "breakfast" is uttered more in the morning and evening than in the day. This is because there are many conversations in the morning and evening saying "today / breakfast / ..." and "tomorrow / breakfast / ...". Also, the utterance “entrance” is uttered frequently from January to May. This is because the number of conversations "entrance / test" and "entrance / ceremony" increases. Such time-varying phenomena cannot be expressed by a language model. Therefore, the bias with respect to the time is introduced as a new evaluation value.

This embodiment will be described with reference to the flowchart of FIG. Assuming that the utterance of this embodiment is performed at about 9:30 (step S10), the corresponding section is “09:00”
１１11: 00 ”. It is searched whether the words“ choice (breakfast) ”and“ choice (lunch) ”are present in the utterance knowledge base 9 (step S20). Here, as shown in FIG. 3, the evaluation values of the words “breakfast” and “lunch” are set in the utterance knowledge base 9 (see FIG. 3), so that “breakfast” is 0.3 and “lunch” Is 0.25
(Step S30). These evaluation values are transferred to the word hypothesis narrowing unit 6 (step S50). The word hypothesis narrowing unit 6 evaluates this evaluation value (weight) for a word that cannot be completely narrowed down using the statistical language model 7.
(Step S60). Referring to the recognition result of FIG. 2, in the word hypothesis "What is breakfast?"
The weighted likelihood is 0.105 obtained by multiplying the output likelihood of 35 and the evaluation value 0.3 output from the speaker state analysis unit 8, and the output of 0.35 in “What is lunch?” 0.0875 obtained by multiplying the likelihood by the evaluation value 0.25 output from the speaker state analysis unit 8 is the likelihood after weighting. Therefore, the output of the word hypothesis narrowing unit 6 can output "what is breakfast" (likelihood: 0.105) having the maximum likelihood after weighting as a recognition result (step S70). In the embodiment according to the second and fourth aspects of the present invention, the input of the speaker state analyzing unit 8 is not a time but other external factors. Here, as an external factor, the loudness of the sound will be described as an input value (FIG. 7).
reference). For example, it is assumed that “turn down the volume” is uttered toward microphone 1 (step S100). As in the case of the first embodiment, the "lower volume" and the "higher volume" may not be distinguished acoustically because the "lower" consonant s may be dropped. In particular, it cannot be distinguished under noise. Therefore, the loudness of the sound is input to the input of the speaker state analysis unit 8, and the control is performed so that the weight of “down” is increased when the sound is loud, and the weight of “up” is increased when the sound is low. The words “raise” and “lower” are searched from the utterance knowledge base 9 (step S200),
If there is a word, an evaluation value is obtained (step S3).
00), and transferred to the word hypothesis narrowing section 6 (step S5).
00), multiplication is performed in the same manner as in claim 1 (step S600), and the maximum value of the likelihood after weighting is output as a recognition result (step S700). Other external factors such as weather, temperature, and clarity can be used as well. For example, the utterance knowledge base 9 when other external factors are clarified is as shown in FIG. This utterance knowledge base 9 is
It is set before recognition, but the contents of the knowledge base can be changed, added and deleted as appropriate.

[0008]

According to the first aspect of the present invention, a word hypothesis that cannot be completely narrowed down even by using a statistical language model is weighted with an evaluation value in consideration of a temporal bias in utterance. More practical narrowing down is possible, and a higher recognition rate can be obtained. According to the second aspect of the present invention, even in an environment where recognition errors are likely to occur, a higher recognition rate can be obtained by introducing an evaluation value in consideration of the sound volume. According to the third aspect of the present invention, for a word hypothesis that cannot be completely narrowed down even by using a statistical language model, weighting is applied to an evaluation value in consideration of a temporal bias in utterance, so that it is more likely to be practical. It is possible to narrow down and obtain a higher recognition rate. According to the fourth aspect of the present invention, even in an environment where recognition errors are likely to occur, a higher recognition rate can be obtained by introducing an evaluation value in consideration of the sound volume.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a voice identification device according to one embodiment of the present invention.

FIG. 2 is a diagram illustrating a result of a voice recognition unit according to the embodiment of the present invention.

FIG. 3 is a diagram illustrating a state of an utterance knowledge base according to an embodiment of the present invention using a temporal factor.

FIG. 4 is a diagram for explaining a state of an utterance knowledge base in an embodiment of the present invention using a loudness as an external factor.

FIG. 5 is a diagram illustrating a state of an utterance knowledge base in the embodiment of the present invention using clarity as an external factor.

FIG. 6 is a flowchart for explaining an embodiment of the present invention using a temporal factor.

FIG. 7 is a flowchart for explaining an embodiment of the present invention using a loudness as an external factor.

[Explanation of symbols]

1 microphone, 2 voice feature extraction unit, 3 voice recognition unit, 4 phoneme HMM, 5 word dictionary, 6 word hypothesis narrowing down unit, 7 statistical language model, 8 speaker state analysis unit, 9 utterance knowledge base.

Claims (4)

[Claims] 1. A speech recognition apparatus that analyzes an input speech into acoustic parameters, compares the speech with a plurality of pre-stored comparison target pattern candidates, and recognizes the speech. Focusing on, means to obtain a value set based on the utterance time, and weighting the likelihood after speech recognition obtained in advance using the obtained value,
Means for narrowing down likely words. 2. A speech recognition apparatus according to claim 1, wherein attention is paid to an external factor, and a likelihood after speech recognition obtained in advance is weighted using a value set for the external factor. A speech recognition device characterized by the following. 3. A speech recognition method for analyzing an input speech into acoustic parameters, comparing the analysis result with a plurality of pattern object candidates stored in advance, and recognizing the speech. Paying attention to the temporal bias of the uttered word, obtaining a value set based on the uttered time, weighting the likelihood after speech recognition previously obtained using the obtained value, A speech recognition method characterized by narrowing down words that are likely to occur. 4. A speech recognition method according to claim 3, wherein attention is paid to an external factor, and a likelihood after speech recognition obtained in advance is weighted using a value set for the external factor. A speech recognition method characterized in that: