JPH10207485A - Speech recognition device and speaker adaptation method - Google Patents

Abstract

(57) [Summary] [Problem] To improve the recognition performance even with a small amount of learning data, further improvement of the performance can be expected if a large amount of data is collected, and the burden on the user for adaptation can be eliminated as much as possible. In a speaker adaptation mode, a phoneme label sequence determination unit (13) obtains correct phoneme sequence information by collating an input voice of a specific speaker with an HMM in a dictionary storage unit (15) corresponding to a correct phoneme sequence. Then, the optimum phoneme sequence information that maximizes the score is obtained by collation with all the HMMs in the dictionary storage unit 15. The adaptation unit 14 learns the average vector and the variance of the phoneme HMM in the dictionary storage unit 15 according to the maximum posterior probability estimation method according to the correct phoneme sequence information, and furthermore, the phoneme label sequence and the optimal phoneme sequence in the correct phoneme sequence information. By comparing with the phoneme label sequence in the information, a speech pattern to which a phoneme label different from the correct phoneme label is assigned is extracted, and the speech pattern is subtracted from the average vector of the phoneme HMM corresponding to the phoneme label.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus having a speaker adaptation function and a speaker adaptation method.

[0002]

2. Description of the Related Art In general, speaker adaptation in speech recognition includes both a method of adapting a dictionary used for recognition to characteristics of a speaker's voice and a method of normalizing a speaker's characteristics included in input speech. Be taken.

First, as a method of adapting a recognition dictionary to a specific speaker's voice, speaker adaptation in a composite similarity method, which is one of statistical recognition methods, is known. The speech recognition dictionary used in the composite similarity method has eigenvalues obtained by principal component analysis of a covariance matrix created from a speech pattern,
It consists of eigenvectors. In order to adapt this voice recognition dictionary to the voice of a specific speaker, a covariance matrix such as the following equation is updated using a voice pattern belonging to each category as a unit of collation, By performing principal component analysis on the updated covariance matrix, a speaker-adapted recognition dictionary is obtained.

K ′ = K + aΣXX ^t where K ′ is a covariance matrix after update, K is a covariance matrix before update, X is a voice pattern belonging to the corresponding category, a is an update coefficient, and t is transpose.

[0005] In this way, by adding the voice uttered by a specific speaker to the covariance matrix created for an unspecified speaker before updating, for example, the recognition reflecting the characteristics of the voice of the specific speaker is performed. You can create a dictionary.

Also, LVQ (Learning Vector Quantiza)
A method called option) has also been proposed. This method
A codebook for generating a code sequence used in a discrete HMM (Hidden Markov Model) is adapted to a speaker. Here, based on the recognition result,
The codebook (code vector) of each category is updated using the input voice pattern. Specifically, when the input voice is determined to be a category different from the correct category, a process of bringing the input voice pattern closer to the correct category and a process of moving the input voice pattern away from the wrong category are performed. (Literature: Shunichi Amari, supervised by Seiichi Nakagawa, Kiyohiro Kano,
Yoichi Higashikura, "Speech / Hearing and Neural Network Model," p.
Further, a speaker adaptation method of a mean vector of a Gaussian distribution in a continuous HMM called a maximum posterior probability estimation method has been proposed. In this method, similarly to the above method, the parameters of the continuous HMM are updated using the voice uttered by the speaker. (Reference: Japanese Patent Application Laid-Open No. 8-95592) On the other hand, a spectrum mapping method has been proposed as a method for normalizing the speaker characteristics of input speech. In the method, a correspondence table for mapping a voice pattern of a specific speaker to a voice pattern of a standard speaker is obtained in advance,
At the time of recognition, the input voice of a specific speaker is converted into the voice of a standard speaker and used for recognition. (Literature: ATR Advanced Technology Series: Automatic translation telephone, pp.70-72, Ohmsha). As a result, without changing the recognition dictionary,
Recognition performance close to that of a standard speaker can be obtained for a specific speaker's voice.

[0007]

As described above, as a speaker adaptation method in speech recognition, a method of adapting a recognition dictionary to a specific speaker's speech by using speech data uttered by a speaker, and a standard speech A method has been proposed in which the input speech of a specific speaker is mapped to the speech of a standard speaker by associating the input speech with a speaker.

However, in the case of a recognition method based on a statistical method, a large number of data to be used for adaptation is required, and a small amount of data has little effect. There is a problem that a great burden is imposed on the speaker. Further, in the maximum posterior probability estimating method, which is effective with a small amount of data, for example, the performance improvement is quickly saturated, so that even if a large amount of voice data can be collected, there is a problem that the recognition rate is not improved to some extent.

Further, in a spectrum mapping method or the like,
In order to perform correct mapping, a large amount of voice data of a specific speaker is required in advance, and the burden on the speaker becomes a problem.

[0010] To reduce the burden on the speaker, a method called unsupervised learning that does not previously assign a correct answer category to the above-mentioned method is being studied. This is a method that enables a dictionary to be learned without the correct answer information for the voice uttered by the speaker. Even if the voice is not uttered in advance for learning, the voice provided for the actual recognition is used. There is an advantage that it can be used for learning as it is, but it may be learned as the wrong category, and compared to supervised learning,
Generally, the degree of improvement in recognition performance is small.

The present invention has been made in consideration of the above problems, and has as its object the purpose of improving the recognition performance even with a small amount of learning data.
It is an object of the present invention to provide a speech recognition apparatus and a speaker adaptation method that can be expected to further improve performance and that can minimize the burden on the user for adaptation.

[0012]

SUMMARY OF THE INVENTION In order to solve the above problem, the present invention relates to a known phoneme sequence corresponding to an input voice of a specific speaker, which is associated with a voice pattern obtained by voice analysis of the input voice. By extracting the correct phoneme sequence information including the information of the matching result by performing the matching with the recognition dictionary (phoneme recognition dictionary) to be performed, and by performing the matching between the voice pattern and all the recognition dictionaries (phoneme recognition dictionary). Phonological sequence information determining means for extracting optimal phonological sequence information including information on a matching result for a phonological sequence giving the maximum likelihood,
According to the correct phoneme sequence information, a first adaptive learning means for learning a corresponding recognition dictionary by a maximum posterior probability estimation method, and comparing the correct phoneme sequence information and the optimal phoneme sequence information to extract a difference part thereof, A second adaptive learning means for learning a corresponding recognition dictionary using the voice pattern in a direction in which the difference is eliminated.

In such a configuration, by providing a known phoneme sequence corresponding to the input speech of a specific speaker, a correct answer including a phoneme section and a phoneme label sequence corresponding to the known phoneme sequence (correct phoneme sequence) is provided. Phoneme sequence information is extracted. Further, by matching the voice pattern of the input voice (input voice pattern) with the entire recognition dictionary, optimal phoneme sequence information including a phoneme section and a phoneme label sequence related to a phoneme sequence having the maximum likelihood (score) is extracted. Here, when the phoneme recognition dictionary is a phoneme HMM, the correct phoneme sequence information and the optimal phoneme sequence information include parameters of a normal distribution including an average vector and a variance for each phoneme, each state, and each mixture (a mixture of normal distributions). Including.

When the correct phoneme sequence information is extracted, the corresponding phoneme recognition dictionary can be learned by the maximum posterior probability estimation method according to the correct phoneme sequence information. In the learning using the maximum posterior probability estimation method, the phoneme recognition dictionary uses the phoneme H
In the case of the MM, the mean vector and the variance, which are parameters of the normal distribution of the HMM, are updated by selectively using the voice pattern of the corresponding phoneme section (having a phoneme label). Normally, the parameters of the normal distribution of each phoneme HMM are present for each state (state number indicating) and each mixture (mixing number indicating), so that the number of normal distributions equal to the number of phonemes × the number of states × the number of mixtures Is learned. However, the parameters of the normal distribution of the phoneme HMM in which no speech pattern of the corresponding phoneme section exists do not belong to the learning.

Next, a difference between the correct phoneme sequence information and the optimal phoneme sequence information is extracted, and learning (updating) of the phoneme recognition dictionary using the input speech pattern is performed in a direction in which the difference is eliminated. Here, the difference is extracted by comparing the phoneme label sequence (correct label sequence) in the correct phoneme sequence information with the phoneme label sequence (optimal label sequence) in the optimum phoneme sequence information. This is a section in the optimal label sequence to which a phoneme label different from the correct phoneme label of the correct label sequence is assigned. By extracting the voice pattern in this section and using it for learning the phoneme recognition dictionary, it is possible to update the phoneme recognition dictionary in a direction in which the difference is eliminated. In particular, when the phoneme recognition dictionary is a phoneme HMM, a speech pattern in a section to which a phoneme label different from the correct label sequence in the optimal label sequence is assigned is converted to a phoneme H of the phoneme of the different phoneme label.
By performing a process of subtracting from the MM average vector, when a pattern similar to the voice pattern appears in the future,
It can be updated so as to be treated as the voice pattern of the correct phoneme label.

As described above, according to the present invention, by using the maximum posterior probability estimation method, an effective speaker adaptation can be realized even when the learning data is small. Further, by comparing the correct label sequence and the optimal label sequence, In addition, since it is possible to perform competitive learning in consideration of the recognition result, it is possible to further improve the recognition performance when a large amount of voice data exists. This also allows the user to maximize the effect of the adaptive function when the amount of learning data is small or large, and as a result, the burden on learning can be greatly reduced.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a basic configuration of a speech recognition device according to one embodiment of the present invention.
The speech recognition device (this device) of FIG. 1 mainly includes a speech input unit 11, a speech analysis unit 12, a phoneme label sequence determination unit 13,
It comprises an adaptation unit 14, a dictionary storage unit 15, a recognition unit 16, a recognized vocabulary storage unit 17, and a control unit 18.

The present apparatus comprises (1) a speaker adaptation mode, and (2)
It is operated in two modes, recognition mode. This mode is selected and designated by the user operating input means such as a keyboard and a mouse. The control unit 18 receives a mode specification from the user, and controls execution of a process in the received mode.

Here, of the two modes, the processing of the (1) speaker adaptation mode, which is a feature of the present invention, will be described with reference to the flowchart of FIG. First, in the speaker adaptation mode, when a user (specific speaker) utters a voice, the voice is A / D-converted by the voice input unit 11 at a predetermined sampling frequency, and is converted into digital time-series signal voice data. The voice input unit 11 applies a fixed time (for example, 8 m) to the voice data thus obtained.
s: This unit is hereinafter referred to as a frame), and the speech power is calculated, and the start and end times of the uttered speech are detected using the time series of the power. Then, the voice input unit 11 extracts the voice data in the start / end section based on the detected start / end time of the voice and sends it to the voice analysis unit 12.

The voice analysis unit 12 performs a frequency analysis on the voice data in the start / end section sent from the voice input unit 11 using, for example, a fast Fourier transform (FFT) or the like to obtain a time-series signal of the voice data. Is converted into time-series data of frequency parameters. Here, a 256-point fast Fourier transform is performed, and the obtained 128-dimensional power spectrum is compressed into a 16-dimensional band-pass filter output by the Bark scale, and a speech pattern including a 16-dimensional feature vector per frame is obtained. Has been generated.

The voice pattern obtained by the voice analysis unit 12, that is, time-series data of frequency parameters as a result of frequency analysis of the input voice is stored in a voice pattern storage unit (not shown).

As described above, when a specific speaker utters m times for n types of words, m voice patterns for each of the n types of words are stored in the voice pattern storage unit.

When a group of voice patterns corresponding to all input voices necessary for the speaker adaptation process is stored in the voice pattern storage unit, the control unit 18 activates the phoneme label sequence determination unit 13. Then, the phoneme label sequence determination unit 13
Performs phoneme label sequence determination processing for obtaining correct phoneme sequence information and optimal phoneme sequence information as described below (step S1).

That is, the phoneme label sequence determining unit 13 determines a known phoneme sequence (correct phoneme sequence) corresponding to the input speech.
Viterbi between the speech pattern obtained by the speech analysis unit 12 and the phoneme HMM (here, the continuous HMM) stored in the dictionary (recognition dictionary) storage unit 15 (corresponding to the correct phoneme sequence). By performing the matching, the section of each phoneme is obtained, and at the same time, which state of each frame in the section corresponds to which state of the phoneme HMM, which mixing degree (normal distribution of which mixing number), and the correct answer including the phoneme label sequence It is stored as phoneme sequence information.

Also, the phoneme label sequence determination unit 13 separately performs the Viterbi comparison between the voice pattern and all the phoneme HMMs stored in the dictionary storage unit 15 without giving a correct phoneme sequence to obtain a maximum score ( The phoneme sequence (likelihood sequence) giving the likelihood), its section (phoneme section), the state of the phoneme HMM for each frame, and the degree of mixing are determined, and the optimal phoneme sequence information including the phoneme label sequence is obtained. Hold as.

Thus, the phoneme label sequence determining unit 13
FIGS. 3 and 4 show examples of the correct phoneme sequence information and the optimal phoneme sequence information obtained by using the case where the input voice is "I" and therefore the phoneme notation is "WATASHI".

The phoneme label sequence determination unit 13 executes the above processing for all the voice patterns obtained by the voice analysis unit 12. When the phoneme label sequence determination processing (step S1) by the phoneme label sequence determination unit 13 ends, control is transferred to the adaptation unit 14. The adaptation unit 14 calculates the above-mentioned 2 obtained by the phoneme label sequence determination unit 13 for each voice pattern.
Using the two phoneme sequence information, the phoneme H
Update (learn) the parameters of the MM.

First, based on the correct phoneme sequence information, the adaptation unit 14 performs the phoneme HMM of each phoneme forming the corresponding correct phoneme sequence.
The parameters (mean vector and variance) of each normal distribution of (continuous HMM) are updated as follows by the maximum posterior probability estimation method as learning targets (step S2).

That is, the adaptation unit 14 controls the phoneme (phoneme category).
The average vector of the normal distribution of the k phoneme HMM is updated by the following equation: μ _k ′ = (αμ _k + Σ _i X _i ) / (α + N), and the variance is similarly calculated by the following equation: σ _k ′ = ｛Σ _i X _i ² − (Α + N) μ _k ′ ² + βσ _k +
Update by αμ _k ^{2 2} / (N + β).

Here, μ _k is an average vector before updating, μ
_k ′ is the average vector after the update, N is the number of audio patterns to be used for learning, X _i is the audio pattern to be used for learning, σ _k is the variance before update, σ _k ′ is the variance after update, and α and β are the update coefficients. , Σ _i X _i is the sum of N voice patterns X _i , Σ _i X
_i ² is the sum of the square values of each of the N voice patterns X _i .

The learning (updating) is performed for each phoneme, each state,
This is performed for all combinations of each mixing degree. Therefore, for example, the number of phonemes is 100, the number of states of each phoneme HMM is 4,
In the case of 5 mixing numbers, 100 × 4 × 5 = 2000 (pieces)
Learning of the normal distribution of. However, if there is no audio pattern for learning the distribution,
The learning is not performed.

The adaptation unit 14 performs the learning (update) described above.
When the processing, that is, the processing (step S2) of learning (updating) the average vector and variance of each normal distribution constituting the phoneme HMM of each phoneme forming the correct phoneme sequence indicated by the correct phoneme sequence information is completed, additional learning of the average vector is performed. Perform processing. In this additional learning process of the average vector, for the sake of simplicity, it is assumed that each phoneme HMM has three states and three degrees of mixture.
This will be described with reference to FIG.

Here, as described above, "I"
Is input as a learning voice. In addition, "WA" which is a phonemic notation of this input voice "I"
As a result of applying the sequence “−TA−SH-I” to the voice pattern of the input voice “I” by Viterbi matching by the phoneme HMM in the phoneme label sequence determination unit 13, the correct phoneme sequence denoted by reference numeral 51 in FIG. It is assumed that a section of each phoneme (W, A, T, A, SH, I) forming the sequence as in the example is obtained (in the process of step S1).

Further, when the phoneme label sequence determining unit 13 finds the optimum phoneme sequence that takes the maximum score for the voice pattern of the input voice “I”, the code 5 in FIG.
2 and the optimal phoneme sequence “WAHTA-SH-I” and each phoneme (W, A,
It is assumed that the section of (H, T, A, SH, I) has been obtained (in the process of step S1).

Here, between the correct phoneme sequence 51 and the optimal phoneme sequence 52, there are six sections in which different phonemes are labeled, and the sections a, b, c, d, e, and
f. Among the sections a to f, for example, the section a is assigned to the phoneme (label) “W” in the correct phoneme sequence 51, but is assigned to “A” in the optimal phoneme sequence 52. The section a includes two frames, a fifth frame and a sixth frame, as is apparent from FIG.

When step S2 is completed, the adaptation unit 14 first sets a frame number j to be processed (indicating the frame position of the frame) to an initial value 1 (step S3).

Next, the adaptation unit 14 determines the j-th frame phoneme label in the correct phoneme sequence 51 (included in the correct phoneme sequence information) and the optimal phoneme sequence 5 (included in the optimum phoneme sequence information).
A comparison is made with the phoneme label of the j-th frame in 2 (step S4). If the two labels are not different (step S5), the adaptation unit 14 advances the frame number j by one (step S6), and then returns to step S4.

In the example of FIG. 3, the phoneme labels of the first to fourth frames of the correct phoneme sequence 51 and the optimal phoneme sequence 52 are:
All are “W” and coincide with each other. On the other hand, the next fifth
The phoneme labels of the frame and the sixth frame, that is, the phoneme labels of the respective frames in the section a are “W” in the correct phoneme sequence 51 and “A” in the optimal phoneme sequence 52, which is different from the above. I have.

When the phoneme label of the j-th frame is different between the correct phoneme sequence 51 and the optimal phoneme sequence 52 as in the example of the fifth frame or the sixth frame (steps S4 and S5). ), The phoneme label name, the HMM state number, and the HMM mixed number of the j-th frame in the optimal phoneme sequence 52 (in the case of the fifth frame in the example of FIG. 3, the phoneme label name = A, the HMM state number = 1, HMM mixing number = 3)
At the same time, the audio pattern of the j-th frame is held (step S7).

Next, the adaptation unit 14 determines whether or not the processing of the last frame has been performed (step S8).
Is advanced by one (step S6), and the process returns to step S4.

In this way, the processing after step S4 is repeated until the processing of the last frame is performed (step S8), and the adaptation unit 14 stores the (correct answer phoneme sequence information) Using the speech patterns of frames whose phoneme labels are different between the correct phoneme sequence 51) and the optimum phoneme sequence 52 (in the optimum phoneme sequence information), the corresponding state number and mixture number in the corresponding phoneme HMM Is updated, and the dictionary storage unit 15 is updated.
(Step S9).

The details of the average vector updating process in step S9 will be described below. Taking the section a as an example, the section a is assigned to “W” in the correct phoneme sequence 51, but is assigned to “A” in the optimal phoneme sequence 52. This section a is a section that should be regarded as “W” instead of “A”.

Therefore, in the present embodiment, when a pattern similar to the voice pattern in the section a appears in the future, the pattern of the section a is set as shown in the following equation so as not to become the pattern of “A”. The voice pattern of the "A" phoneme HM
A process of subtracting from the average vector of M (average vector update process) is performed.

Μ _k ″ = μ _k ′ + (γ / N) { _i (X _i −μ _k ′)} where μ _k ′ is the average vector before update, and μ _k ″ is the average vector after update. , Γ are update coefficients (negative values), X _i is a voice pattern to be used for learning, N is the number of voice patterns to be used for learning, k is a phoneme category, and Σ _i (X _i −μ _k ′) is N
Represents the sum of X _i −μ _k ′ for a plurality of voice patterns X _i .

The phoneme H whose average vector has been updated in this way
The MM is stored again in the dictionary storage unit 15 as described above, and is subjected to a recognition process. As described above, each phoneme HMM
By updating (learning) the average vector and the variance of the specific speaker using the voice uttered by the specific speaker, the phoneme HMM can be adapted to the speaker, and the recognition performance can be improved. The result of confirming the improvement of the recognition performance in the case where this parameter updating (learning) method (speaker adaptation method) of the phonological HMM is applied to a speech recognition apparatus in a 500 word recognition experiment (three males) was obtained. The average value of the recognition rate for each specific speaker with respect to the number of voice data to be provided) is shown by a solid line in FIG. For reference, the case where only the maximum a posteriori probability estimation method is used is indicated by a broken line. In FIG. 5, the horizontal axis is the number of audio data to be used for learning, and the vertical axis is the recognition rate.

As is apparent from FIG. 5, the method applied in the present embodiment shows higher recognition performance even when the learning data is small, and also has a higher learning data than the case where only the maximum posterior probability estimation method is used. There is no saturation of the recognition performance as the number increases.

Now, as a result of updating the HMM parameters (mean vector and variance therein) according to the flowchart (steps S1 to S9 in FIG. 2), for example, if the recognition performance has increased by a predetermined ratio or more, a series of speaker adaptations The process ends (step S10). On the other hand, if the rate of increase in the recognition performance has not reached the predetermined ratio, the processing after step S1 is performed again. Note that as the condition for terminating the speaker adaptation process, the number of processes (the number of times the processes of steps S1 to S9 are repeated) may be used instead of the rate of increase in recognition performance.

Next, the processing in the recognition mode will be described. Note that the recognition mode is not directly related to the present invention. Therefore, here, a brief description will be given of a generally performed recognition process as an example.

The processing of the speech input unit 11 and the speech analysis unit 12 in the recognition mode is the same as that of the above-described speaker adaptation mode, and the speech analysis unit 12 acquires a speech pattern representing the characteristics of the input speech.

The voice pattern of the input voice obtained by the voice analysis unit 12 is sent to the recognition unit 16. Recognition unit 16
For each vocabulary stored in the recognized vocabulary storage unit 17, Viterbi matching with a voice pattern is performed using the phoneme HMM in the dictionary storage unit 15 to obtain a score (likelihood). Here, for example, when the recognized vocabulary is a word, the recognizing unit 16 forms a word HMM by concatenating the corresponding phoneme HMMs according to the phoneme sequence forming the word, and performs matching with the voice pattern for each word HMM. . After obtaining the scores for all the vocabularies in this way, the recognition unit 16 outputs the vocabulary having the highest score as the recognition result.

The function of each part of the speech recognition apparatus having the configuration shown in FIG. 1 described above is the same as that of the personal computer shown in FIG. 7 having a speech input function in which a computer, for example, a built-in microphone is incorporated or a microphone input terminal is provided. The computer 70 is mainly composed of a speech analysis unit 12, a phoneme label sequence determination unit 13, an adaptation unit 14, a recognition unit 16, and a control unit 1.
A recording medium, such as a floppy disk (FD) 71, on which a program for functioning as the computer 8 is recorded, is used.
0, and the program recorded on the floppy disk 71 is read and executed by the personal computer 70.

The speech analysis conditions described in the above embodiment and the phoneme sequence shown in FIG. 5 are merely examples, and the present invention is not limited to these contents. In addition, the present invention is not limited to the above-described embodiment, and can be implemented with various modifications without departing from the gist thereof.

[0053]

As described in detail above, according to the present invention, since the maximum posterior probability estimation method is used, even when the learning data is small, the effect of speaker adaptation is remarkable, and the optimal phoneme is further improved. Since the competitive learning of the sequence and the correct phoneme sequence is used together, even if the learning data increases, the convergence does not converge and further improvement of the recognition performance can be expected as compared with the case of using only the maximum posterior probability estimation method. In addition, this allows the user to control the utterance of the speaker adaptation voice data depending on the situation without uttering a large amount of speech at a time, so that the user's burden on adaptation can be reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a basic configuration of a speech recognition device according to an embodiment of the present invention.

FIG. 2 is an exemplary flowchart for explaining processing in a speaker adaptation mode according to the embodiment;

FIG. 3 is a view showing a part of an example of correct phoneme sequence information and optimum phoneme sequence information obtained by a phoneme label sequence determination unit 13 in FIG. 1;

FIG. 4 is a view showing another part of an example of correct phoneme sequence information and optimum phoneme sequence information obtained by the phoneme label sequence determination unit 13 in FIG. 1;

FIG. 5 is a diagram showing a section in which different phoneme labels are assigned to a correct phoneme sequence and an optimal phoneme sequence.

FIG. 6 is an exemplary view showing the effect of the speaker adaptation method applied in the embodiment in comparison with the case where only the maximum posterior probability estimation method is used;

FIG. 7 is an exemplary external view of a personal computer that implements the voice recognition device of FIG. 1;

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 11 ... Speech input part 12 ... Speech analysis part 13 ... Phoneme label sequence determination part (phoneme sequence information determination means) 14 ... Adaptation part (first adaptive learning means, second adaptive learning means) 15 ... Dictionary storage part 16 ... Recognition unit 17: Recognized vocabulary storage unit 18: Control unit

Claims (4)

[Claims] 1. A voice input unit for inputting a uttered voice, a voice analysis unit for analyzing a voice input by the voice input unit to obtain a voice pattern representing the feature, and a matching unit for each phoneme. Dictionary storage means for storing a group of recognition dictionaries; and recognition means for executing recognition processing of a voice pattern obtained by the voice analysis means using a recognition dictionary in the dictionary storage means in a recognition mode. In the speaker recognition mode, in a speaker adaptation mode, for a known phoneme sequence corresponding to an input speech of a specific speaker, a correspondence between a speech pattern obtained from the input speech by the speech analysis unit and the dictionary storage unit. The correct phonological sequence information including the information of the matching result is extracted by performing the matching with the recognition dictionary to be recognized, and at the same time, the voice pattern and all the recognitions in the dictionary storage unit are recognized. A phonological sequence information determining means for extracting optimal phonological sequence information including information on a matching result relating to a phonological sequence giving the maximum likelihood by performing a check with a dictionary; and the correct phoneme extracted by the phonological sequence information determining means. A first adaptive learning means for learning a corresponding recognition dictionary in the dictionary storage means by a maximum a posteriori probability estimating method according to the sequence information; and the correct phoneme sequence information extracted by the phoneme sequence information determining means and the optimal The phoneme sequence information is compared to extract the difference, and learning of the corresponding recognition dictionary in the dictionary storage is performed using the voice pattern obtained by the voice analyzer in the direction in which the difference is eliminated. A speech recognition device comprising: a second adaptive learning unit. 2. A speech input means for inputting uttered speech, a speech analysis means for analyzing a speech inputted by the speech input means to obtain a speech pattern representing the feature, and a collation for each phoneme. Dictionary storage means for storing a group of phoneme HMMs; and recognition means for executing recognition processing of the voice pattern obtained by the voice analysis means using the phoneme HMMs in the dictionary storage means in a recognition mode. In the speaker recognition mode, in a speaker adaptation mode, for a known phoneme sequence corresponding to an input speech of a specific speaker, a correspondence between a speech pattern obtained from the input speech by the speech analysis unit and the dictionary storage unit. The correct phoneme sequence information including the phoneme label sequence is extracted by performing the matching with the phoneme HMM to be performed. Phonological HM of
A phoneme sequence information determining means for extracting optimal phoneme sequence information including a phoneme label sequence relating to a phoneme sequence giving the maximum likelihood by performing matching with M; and the correct phoneme sequence extracted by the phoneme sequence information determining means. First adaptive learning means for learning the parameters of the corresponding phoneme HMM in the dictionary storage means according to the maximum posterior probability estimation method according to the information, and the correct phoneme sequence information extracted by the phoneme sequence information determination means. A phoneme label sequence is compared with a phoneme label sequence in the optimal phoneme sequence information, and a speech pattern to which a phoneme label different from the correct phoneme label is assigned is extracted from the speech patterns obtained by the speech analysis means. Then, using the voice pattern, the corresponding HMM is stored in the dictionary storage means in the direction in which the difference between the phoneme labels is eliminated. Speech recognition apparatus characterized by comprising a second adaptive learning means for performing learning of the meter. 3. A voice input means for inputting uttered voice, a voice analysis means for analyzing a voice input by the voice input means to obtain a voice pattern representing a feature thereof, and a normal means comprising an average vector and a variance. A dictionary storing means for storing a group of phonemic HMMs including distribution parameters, and a recognition process of a voice pattern obtained by the voice analyzing means in a recognition mode, using a phonemic HMM in the dictionary storing section. And a recognition unit for executing, in a speaker adaptation mode, for a known phoneme sequence corresponding to an input voice of a specific speaker, a voice pattern obtained from the input voice by the voice analysis unit and the voice pattern. The correct phoneme sequence information including the phoneme label sequence is extracted by matching with the corresponding phoneme HMM in the dictionary storage means, and the speech pattern is extracted. Ting all phonemes HM of the dictionary storage in means
A phoneme sequence information determining means for extracting optimal phoneme sequence information including a phoneme label sequence relating to a phoneme sequence giving the maximum likelihood by performing matching with M; and the correct phoneme sequence extracted by the phoneme sequence information determining means. First adaptive learning means for learning the average vector and variance of the corresponding phoneme HMM in the dictionary storage means by a maximum a posteriori probability estimating method according to the information; and the correct phoneme extracted by the phoneme sequence information determining means. The phoneme label sequence in the sequence information is compared with the phoneme label sequence in the optimal phoneme sequence information, and a phoneme label different from the correct phoneme label is assigned from the speech patterns obtained by the speech analysis means. A voice pattern is extracted, and the voice pattern is extracted from the average vector of the phoneme HMM in the dictionary storage means corresponding to the phoneme label. Speech recognition apparatus characterized by comprising a second adaptive learning means for pulling. 4. A group of recognition dictionaries for each phoneme stored in the dictionary storage means for recognition processing of a voice pattern representing characteristics of the input voice obtained by analyzing the input voice. A speaker adaptation method for learning so as to adapt to a known phoneme sequence corresponding to an input speech of a specific speaker, by comparing a speech pattern of the input speech with a corresponding recognition dictionary in the dictionary storage means. , The correct phoneme sequence information including the information of the matching result is extracted, and the voice pattern is checked against all the recognition dictionaries in the dictionary storage means, so that the matching of the phoneme sequence giving the maximum likelihood is performed. A first step of extracting optimal phoneme sequence information including result information, and the dictionary storage method using a maximum posterior probability estimation method according to the correct phoneme sequence information extracted in the first step. And a second step of learning a corresponding recognition dictionary in the first step, and comparing the correct phoneme sequence information and the optimal phoneme sequence information extracted in the first step to extract a difference part thereof. A third step of learning a corresponding recognition dictionary in the dictionary storage unit using the voice pattern in the direction to be canceled.