JPH10214097A - Adapting method for speech feature quantity, speech recognition device, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To adapt the feature quantity of a speech to be recognized in real time. SOLUTION: Feature vectors of respective frames are extracted from an inputted object speech and stored as adapted data in a feature quantity storage part 21. A mean feature quantity calculation part 22 calculates the mean of the adapted data from feature vectors of the stored adapted data. A mean distribution value calculation part 25, on the other hand, calculates the mean distribution value of recognition models corresponding to phonemes in the adapted data. A difference value calculation part 23 calculates the difference value between the mean of the said distribution mean values and the mean of feature quantities of the adapted data. The difference value is updated each time the adapted data change. A feature quantity normalization part 13 calculates a normalized feature quantity by subtracting the said difference value from the feature vector of the object speech and a recognition processing part 14 performs a recognition processing by using the normalized feature quantity.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for adapting speech recognition, and more particularly to a technique for quickly correcting the degradation of recognition performance when distorted speech is input.

[0002]

2. Description of the Related Art Generally, a speech recognition apparatus prepares a subword model learned in advance from a large number of speech models and generates a word model by connecting a plurality of subword models when registering a word. The generated word model is learned as needed, and is used in the recognition process. At the time of recognition, for example, after converting an analog uttered voice input from a microphone into a digital signal, a feature amount (feature vector) of the digital signal is calculated, and the feature amount is compared with a word model. Then, as a result of the collation, the most similar word model is output as a recognition result.

[0003] By the way, the feature amount of a voice differs depending on the speaker, and even for the same speaker, the feature amount may vary depending on the situation at the time of speech, physical condition, and the like. Further, for example, in the case of speech input via a telephone line, multiplication noise is mixed in the speech and distortion occurs. as a result,
Originally, even voices having the same feature amount may be recognized as different voices. That is, the performance of speech recognition is degraded. For this reason, in recent years, the use of a probability distribution model to adapt a feature amount of speech has been increasing.

A conventional adaptation method of this kind will be described below with reference to FIGS.
This will be described with reference to FIG. First, the basic configuration of a conventional speech recognition device will be described with reference to FIG. This voice recognition device 2
A voice input unit 71 for inputting a target voice captured by a microphone or the like, that is, a voice to be subjected to voice recognition;
A feature amount extraction unit 72 for extracting a feature amount (feature vector) of the input speech, a feature amount normalization unit 73 for normalizing the extracted speech feature amount, and recognition based on the normalized speech feature amount A recognition processing unit 74 for performing processing and a recognition result output unit 75 for outputting a recognition result are arranged in this order. Further, in order to realize the above-described adaptation processing, the feature amount extraction unit 72
A feature amount accumulating unit 81 that accumulates the feature amounts extracted in the entire utterance and an average calculating unit 82 that calculates an average of the feature amounts of the entire utterance are provided. Further, the recognition processing unit 74
In order to prepare the above-mentioned word model to be collated, a recognized word registration unit 83 for inputting a word, a word model creation unit 84 for creating a word model corresponding to the input word or the word to be recognized, A recognition word dictionary unit 85 in which models are learned and stored as needed, and a word model storage unit 86 for storing a plurality of word models used directly by the recognition processing unit 74 are provided.

[0005] The procedure of speech recognition in the speech recognition apparatus 2 is as shown in FIG. First, the feature amount of the target voice input from the voice input unit 71 is extracted for each frame by the feature amount extracting unit 72 (step S201). The extracted feature amount is used as a feature amount extraction unit 8 from the beginning to the end of the utterance.
1 (step S202), and the average calculator 82 calculates the average of the feature amounts of the entire utterance (step S203). The feature amount normalization unit 73 normalizes the feature amount by subtracting the average of the feature amounts of the entire utterance from the feature amount of each frame (step S204). As a result, the components of the frequency characteristics (transfer characteristics of the analog device) of the recording system in which the word model and the target voice are recorded are subtracted, and the distortion of the feature amount of the target voice is removed. The recognition processing unit 74 performs a speech recognition process based on the normalized feature amount, that is, performs collation with the word model stored in the word model storage unit 86 (step S205). The above processing is performed at the time of learning and at the time of recognition.

FIG. 8 shows how feature values of a frame i are normalized in the case of an utterance composed of N frames. The recognition result of the recognition processing unit 74 is output to the recognition result output unit 7.
After that, it is sent to the subsequent processing means.

The above-described adaptation method uses cepstrum mean subtraction (CMS: Cepstrum Mean Subtraction or CMS).
It is called MN (Cepstrum Mean Normalization) and is often used in conventional speech recognition devices.

[0008]

As described above, conventionally, the feature amount of the target speech is adapted by subtracting the average of the feature amounts of the entire utterance from the feature amount extracted for each frame. . In such an adaptation method, although the distortion of the target voice can be removed and the recognition performance can be improved, it is indispensable to calculate the average of the feature amounts of the entire utterance. For this reason, there is a problem that the adaptation process cannot be started until the utterance ends, and it is not possible to cope with the quick recognition process.

On the other hand, one of the methods for solving this problem is Q
& A format. In the Q & A format, for example, if you want to enter an address, "where is the prefecture?", "Is Tokyo", and where is the ward or municipal name? , "Setagaya-ku", ..., the question from the device side and the answer from the speaker (input of the target voice) are repeated. In such a Q & A format, the average (vector) of the feature amount of the immediately preceding target voice is obtained, and the average is subtracted from the feature value of each frame of the current target voice, so that quick adaptation processing is possible. . However, in the case of the Q & A format, since the voice is adapted to the voice uttered in advance, the adapted data (prior voice) and the phoneme (phoneme) included in the target voice are different. For this reason, although a certain improvement in performance can be seen in comparison with the case where no adaptation processing is performed, the recognition performance is significantly deteriorated as compared with the above-mentioned CMS or the like.

[0010] Therefore, an object of the present invention is to provide a speech recognition apparatus which can prevent a deterioration in recognition performance when a target speech is distorted, and enables a rapid adaptation processing of a feature used for speech recognition. Is to provide.

[0011]

SUMMARY OF THE INVENTION The present invention first provides an improved method for adapting features of a target speech input to a speech recognition device. In the adaptation method, an average of a feature amount of the target voice in a predetermined voice section is calculated, and an average of a distribution average value of a recognition model corresponding to a phoneme included in the target voice in the voice section is calculated. And a step of normalizing the audio feature amount of the audio section in accordance with the difference value between the averages.

In the method of the present invention, preferably, for example, a mixed continuous hidden Markov model (hereinafter, referred to as HMM) is used as the recognition model, and Viterbi from the target voice is used as the voice section.
A speech section corresponding to the latest n phonemes obtained by the algorithm is used. The voice section is updated every time the latest phoneme is obtained by the Viterbi algorithm, and the difference value is also updated for each voice section.

[0013] The present invention also provides a speech feature amount extraction unit for extracting a speech feature amount from an input target speech, and a recognition processing unit for adapting the extracted speech feature amount and recognizing the target speech. And a speech recognition device having the following. In this voice recognition device, the recognition processing means includes:
First means for calculating an average of the feature amount of the target voice for each predetermined voice section, second means for obtaining a recognition model corresponding to a phoneme string included in the target voice in the voice section, distribution average of the obtained recognition model A third means for calculating an average of the values, and a fourth means for normalizing the audio feature amount of the audio section in accordance with a difference value between the calculated average of the characteristic amounts and the average of the distribution average value. The recognition processing is performed using the normalized speech feature amount updated every time.

[0014] The present invention further provides a recording medium for realizing the above adaptation method on a computer device.
The recording medium is a computer-readable recording medium that records a program for causing a computer device to function as a voice recognition device and adapting a feature amount of an input target voice, and the program is stored in a computer device. At the time of reading, at least the following processing is executed. (1) a method of adapting the feature amount of the target voice, wherein the process of calculating an average of the feature amount of the target input target voice in a predetermined voice section; (2) a process included in the target voice of the voice section (3) a process of averaging the distribution average value of the recognition model corresponding to the phoneme to be performed, and (3) a process of normalizing the feature amount of the speech section according to a difference value between the average of the feature amount and the average of the distribution average value.

[0015]

Next, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram of a speech recognition device 1 according to an embodiment of the present invention. The speech recognition device 1 is realized by a personal computer or a workstation, and is formed by reading and executing a predetermined program by a personal computer or the like, and includes a speech input unit 11, a feature amount extraction unit 12, a feature amount. Normalization unit 13, recognition processing unit 14, recognition result output unit 1
5, feature amount storage unit 21, feature amount average calculation unit 22, difference value calculation unit 23, phoneme string storage unit 24, distribution average value average calculation unit 25, recognition word registration unit 26, word model creation unit 2
7. It has functional blocks of a recognized word dictionary unit 28 and a word model storage unit 29.

The above-mentioned program is usually recorded in a storage device installed inside or outside a personal computer or the like, and is distributed integrally with the personal computer or the like. More specifically, the present invention can be implemented by a program that can realize each of the functional blocks in a personal computer or the like, and the recording form or distribution form is not necessarily limited to the above-described case. For example, flexible disk or compact disk type R
It is recorded in a computer-readable form on a portable recording medium such as an OM or a digital video disk, or on a program server on a private network accessible by a personal computer or the like. The data may be read to an external storage device.

The contents of each functional block in the speech recognition apparatus 1 are as follows. The feature amount of the target speech input from the speech input unit 11 is extracted by the feature amount extraction unit 12, normalized by the feature amount normalization unit 13, and then subjected to speech recognition by the recognition processing unit 14, and the result is recognized. The point output from the result output unit 15 is basically the same as that of the conventional device 2 shown in FIG. Further, the feature amount extracted by the feature amount extraction unit 12 is stored in the feature amount storage unit 21, a recognition word registration unit 26, a word model creation unit 27, a recognition word dictionary unit 28, and a word model storage unit. 29 are basically the same as the conventional device 2.

The features of the speech recognition apparatus 1 of the present embodiment include, in addition to the same functional blocks as those of the above-described conventional apparatus, an average calculating section 22 for calculating an average of the features of the target voice for each predetermined voice section. A phoneme string storage unit 24 for storing phonemes recognized by the recognition processing unit 14; and a distribution average value for finding a recognition model corresponding to the recognized phoneme from the recognition word dictionary unit 28 and calculating an average of the distribution average value. Average calculator 25
And a difference value calculating unit 23 for calculating a difference value of each average, and the feature amount normalizing unit 13 performs normalization processing based on the calculation result by the average difference value calculating unit 23.

Hereinafter, the operation of the speech recognition apparatus 1 according to the present embodiment will be specifically described. Here, a Viterbi algorithm is used as a phoneme string recognition format, and HM is used as a recognition model.
M is used. First, a general theory of the recognition process using the HMM and the Viterbi algorithm will be described.

Now, one model (HMM) obtained by connecting HMMs (M1, M2) respectively corresponding to two phonemes w1, w2 is defined as M3. Assuming that the output symbol sequence corresponding to the target voice is a code of y1, y2, y3,... YT, a model M of the code y1, y2, y3,.
Likelihood P (y1, y2, y3,... YT | M3)
Can be approximated by the following equation (1).

[0021]

(Equation 1)

From the above equation, the boundary t of Viterbi alignment, that is, the boundary of phonemes can be known. HMM
In speech recognition likelihood calculation by
When a category indicating the maximum value of the above approximate expression is output as a recognition result instead of the Welch score, baum-
It has been reported by papers and the like that the performance is equivalent to that when the weld score is used. In this case, one of the advantages generated by using the Viterbi algorithm is that the calculation efficiency is improved.

Next, a method for adapting a speech feature in the speech recognition apparatus 1 will be described in detail with reference to FIGS. Here, as the target voice, “red (aka
i) "is input. In addition, as a feature amount of the target voice, a known cepstrum, △ cepstrum,
△ Use power.

In FIG. 2, first, the input target voice “red (akai)” is segmented into a predetermined frame length, and a feature amount (feature vector) of each frame is extracted (step S101). Specifically, a cepstrum cep (j) for each frame is extracted. j is a frame number. Further, a cepstrum cep (j) in a certain voice section is used as adaptation data, and a recognition model corresponding to a phoneme included in the adaptation data is extracted from the recognition word dictionary unit 28 (step S102). Then, the average of the distribution average value of the extracted recognition model is calculated (step S103). An example of how to calculate the average of the distribution average value of the recognition model is as shown in FIG. 3. When the corresponding recognition model is the recognition model A, B, C (A vector, B vector, C vector), An average vector (A + B + C vector) is obtained. On the other hand, the average of the feature amount is calculated from the adaptation data (step S104). This average is the average of the cepstrum (j) in the thick line portion in FIG. Here, the speech section of the two latest phonemes output by the Viterbi algorithm is used. However, the number of voice sections may be other than two and can be arbitrarily set.

Next, a difference value dif (j) of each average calculated in steps S103 and S104 is calculated (step S105). Thereafter, the difference value dif (j) is subtracted from the cepstrum cep (j) to calculate a normalized cepstrum nsep (j), and the cepstrum and △ power are calculated and recognized from the normalized cepstrum nsep (j). (Step 10)
6). The recognition processing unit 14 performs a recognition process based on the calculated normalized cepstrum ncep (j), △ cepstrum, and △ power, and extracts a word model having an optimum likelihood from the word model storage unit 29 (step S10).
7).

Next, a method for calculating the difference value (j) and a method for updating the difference value (j) will be described. It is assumed that the initial value dif0 of the difference value (j) is set in advance. When less than two voices are output by the Viterbi algorithm, for example, in the above example, “red (aka
When only / a / of "i)" is output, the above initial value dif0 is used as the difference value (j). 2 phonemes
When the number is output, that is, when / a / and / k / are output, the difference value (j) (= dif0) is updated. The equation for calculating the difference value dif (j) is as shown in the following equation (2).

[0027]

(Equation 2)

Here, dif (j-1) represents a difference value of the immediately preceding voice section. Further, t1 is the number of frames in the voice section corresponding to the phoneme output first of the two phonemes, and t2 is the frame number of the voice section corresponding to the phoneme output later in the two phonemes. t is the number of the current frame, N is the number of Gaussian mixtures in the HMM, I
Is the number of states in the HMM, μ1 is the distribution average of the HMM corresponding to the previously output phoneme of the two phonemes, and μ2 is the distribution average of the HMM corresponding to the second output phoneme of the two phonemes. . Thereafter, each time a new phoneme is output by the Viterbi algorithm, the difference value dif (j) is updated.

FIG. 6 shows that / a
/, Number of frames corresponding to / k / (= voice section ta + t
FIG. 9 is an explanatory diagram showing how to find an element of the difference value (j) when k) is “13”. In the illustrated case, the voice section ta corresponding to the previously output phoneme / a / is "7", and the voice section tk corresponding to the phoneme / k / output later is "6".
It is. D1 in this case, that is, the average of the feature amounts of the adaptation data is expressed by the following equation (3).

[0030]

(Equation 3)

The average D2 of the distribution average value for the previously output phoneme / a / is expressed by Equation 4, and the average D3 of the distribution average value for the phoneme / k / output later is expressed by Equation 5.
It is expressed by an equation.

[0032]

(Equation 4)

[0033]

(Equation 5)

As described above, according to the present embodiment, the characteristic of the target speech is determined according to the difference between the average of the feature amounts of the adaptation data based on the target speech and the average of the distribution average of the recognition model corresponding to the phoneme. Since the amount is normalized and the difference value is updated for each voice section that changes at any time, real-time adaptation is possible.

Therefore, even if the target voice is distorted, the deterioration of the recognition performance is quickly corrected. Thus, for example, in an application for recognizing a target voice input via a telephone line, a normal wired telephone, a cordless telephone, an analog mobile phone, a digital mobile phone, a PHS
Thus, the influence of the difference in the telephone network and the influence of the difference in the telephone can be reduced, and the recognition performance can be improved.

[0036]

As is apparent from the above description, according to the present invention, it is possible to prevent the recognition performance from deteriorating when the target speech is distorted, and to quickly adapt a feature used for speech recognition. An enabling approach can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating a procedure of a recognition process according to the embodiment;

FIG. 3 is an explanatory diagram showing an example of a method of obtaining an average of distribution average values according to the embodiment.

FIG. 4 is an explanatory diagram of adaptive data and a voice section used in the embodiment.

FIG. 5 is an explanatory diagram showing the update timing of the difference value according to the embodiment.

FIG. 6 is an explanatory diagram showing a specific example of calculating a difference value according to the embodiment;

FIG. 7 is a block diagram of a conventional voice recognition device.

FIG. 8 is an explanatory diagram of a conventional adaptation method.

FIG. 9 is a diagram illustrating a procedure of a conventional voice recognition process.

[Explanation of symbols]

 1, 2 voice recognition device 11, 71 voice input unit 12, 72 feature value extraction unit 13, 73 feature value normalization unit 14, 74 recognition processing unit 15, 75 recognition result output unit 21, 81 feature value storage unit 22 feature value Calculation unit 23 difference value calculation unit 24 phoneme string accumulation unit 25 average calculation unit of distribution average value 26,83 recognized word registration unit 27,84 word model creation unit 28,85 recognized word dictionary unit 29,86 word model storage unit

Claims (6)

[Claims] 1. A method for adapting a feature amount of a target voice input to a voice recognition device, comprising: calculating an average of a feature amount of the target voice in a predetermined voice section; Calculating an average of the distribution average values of the recognition models corresponding to the phonemes included, and normalizing the characteristic amounts of the speech section according to a difference value between the calculated average of the characteristic amounts and the average of the distribution average values. A method of adapting a speech feature quantity, characterized by including: 2. The method according to claim 1, wherein the recognition model is a mixed continuous hidden Markov model, and the speech section is a speech section corresponding to the latest n phonemes obtained from the target speech by a Viterbi algorithm. 2. The adaptation method according to claim 1, wherein: 3. The adaptation method according to claim 2, wherein the voice section is updated each time a latest phoneme is obtained by the Viterbi algorithm. 4. The adaptation method according to claim 1, wherein the difference value is updated for each voice section. 5. A speech feature amount extraction unit for extracting a speech feature amount from an input target speech, and a recognition processing unit for adapting the extracted speech feature amount to perform recognition processing of the target speech. A first means for calculating an average of feature amounts of a target voice for each predetermined voice section, a second means for obtaining a recognition model corresponding to a phoneme sequence included in the target voice in the voice section, A third means for calculating an average of the distribution average value of the acquired recognition model; and a third means for normalizing the audio feature value of the voice section in accordance with a difference value between the calculated average of the feature value and the average of the distribution average value. A speech recognition apparatus including four means, wherein the recognition processing is performed using the normalized speech feature amount updated for each speech section. 6. A computer-readable recording medium that records a program for causing a computer device to function as a speech recognition device and adapting a feature amount of an input target speech, wherein the program includes: A method for adapting a feature amount of a target voice, the process of calculating an average of a feature amount of the target input target voice in a predetermined voice section, a recognition model corresponding to a phoneme included in the target voice of the voice section. The computer device executes a process of averaging the distribution average value of the above, and a process of normalizing the feature amount of the voice section in accordance with a difference value between the average of the feature amount and the average of the distribution average value. A recording medium characterized by the above-mentioned.