JPH1097273A - Speaker adapting method for voice model, speech recognizing method using the same method, and recording medium recorded with the same method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase the recognition rate by extracting model series of voice models for unspecified speakers by using the voice of a speaker to be recognized and adapting the speech models for unspecified speakers by the model series. SOLUTION: A speech model storage part 12 is previously stored with the speech models for unspecified speakers corresponding to a recognition category as a reference model dictionary. A procedure of implementing a speech recognizing method is previously recorded in a storage part 10M in a control part 10, which controls the processes of respective parts 11 to 16 of the speech recognition part by following the procedure. Then model series are so extracted that a correct model series is included with high probability, and speaker adaption to the respective model series is performed; likelihood to the object speech to be recognized when speech models after the speaker adaptation are used is compared and the speech model having the highest value is selected to perform adaptation based upon a more correct model series.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the use of speech of a large number of speakers to identify features of speech corresponding to recognition categories such as phonemes and words, for example, using a hidden Markov model (Hidden Marko model).
v Model, hereafter referred to as HMM) using the speaker-independent speech model
The present invention relates to a speaker adaptation method for a speech model adapted to increase the recognition rate for the speaker, a speech recognition method using the method, and a recording medium recording the method.

[0002]

2. Description of the Related Art In a speech recognition method, a feature parameter sequence of an input speech is compared with a reference speech (word or phoneme) model sequence created in advance, and the most probable speech model sequence is output as a recognition result. A reference model dictionary, which is a set of such reference speech models used for speech recognition, is created as follows, for example. For example, speech of many speakers is subjected to, for example, linear prediction analysis to obtain a feature parameter sequence. A reference word model or phoneme model (hereinafter collectively referred to as a speech model) is created for each word or phoneme from the feature parameter sequence, and is used as a reference model dictionary.

[0003] In recent years, in speech recognition, as a reference speech model, an HMM representing a speech by a statistic of a feature parameter sequence such as a cepstrum obtained for every certain frame, for example.
In many cases, the HMM will be briefly described with reference to FIG. For example, in a phoneme HMM, a feature parameter sequence corresponding to each phoneme portion generally has a feature divided into a start region, an intermediate region, and an end region. Therefore, as shown in FIG. It can be specified by state. However, the boundary of each area is not clear because the sound phenomenon is continuous,
In A, boundaries of each area are clearly shown for simplicity of explanation. These areas are called first, second, and third states, and S
To be represented by _1, S _2, S _3.

[0004] Each state of the phoneme HMM represents the distribution of a feature parameter sequence corresponding to each region of the phoneme, and is actually represented by a combination of Gaussian distributions in the dimension of the feature parameter vector. FIG. 1B is an example in which each state is represented by a combination of four Gaussian distributions when the dimension of the feature parameter vector is one-dimensional. At that time, the state S _i is the average value of the four Gaussian distributions m _i = {m _i1 , m _i2 , m _i3 , m _i4 }, and the variance σ _i = {σ
_i1 , _σi2 , _σi3 , _σi4 }, and weighting factors w _i = {w _i1 , w _i2 ,
w _i3 , w _i4 }. Such a distribution represented by a combination of a plurality of Gaussian distributions is called a Gaussian mixture distribution.

Further, in each of the first, second, and third states S ₁ , S ₂ , and S ₃ , the probability a ₁₁ , a ₂₂ , and a ₃₃ of transition to the same state and transition to the next state are set. Probability a ₁₂ , a ₂₃
Is defined as shown in FIG. 1C, for example. As described above, a set of statistical parameters of the average value, variance, weighting factor, and state transition probability of all Gaussian mixture distributions of all states is represented by an HMM representing this phoneme
And is represented by θ. The reference model dictionary stores model parameters of the speech HMM for all predetermined phonemes (or words).

When speech recognition is performed, when a feature parameter sequence obtained from input speech is applied to a possible speech HMM sequence, the best-fit speech HMM sequence is output as a recognition result. Note that the degree of the fit is called the likelihood of the HMM, and in fact, the probability calculated from the Gaussian mixture distribution when the feature parameter sequence is stochastically allocated to each state included in the speech HMM sequence, and the state transition Obtained as a product of probability. In order to adapt the speech model to the speech of an unspecified speaker, which will be described later, the model parameters θ (for example, three parameters) of each speech HMM are set so that the likelihood in the speech model sequence corresponding to the speech of the speaker is maximized. The mean value of the Gaussian mixture m ₁ , m ₂ ,
m ₃ ).

[0007] For example, an interactive automatic ticketing apparatus using voice recognition installed in a station is required to recognize the voice of the destination station name uttered by an unspecified user and issue a corresponding ticket. You. In order to increase the recognition rate of the unspecified speaker's voice, the reference model dictionary is adapted to the user using the voice uttered by the user, and the user is used to adapt the reference model dictionary. Re-recognition of voice is conceivable.

In general, this speaker adaptation technique can be considered in two cases: a case where the utterance content of voice data used for the adaptation is known (with teacher) and a case where it is unknown (without teacher). In addition, this technology can be classified into an offline type, in which the recognition system collects an appropriate amount of voice data in advance and uses it for adaptation, and an online type, in which unsupervised adaptation is performed by using the utterance at each recognition. .

Unsupervised online speaker adaptation is referred to as immediate speaker adaptation, in which many users are replaced, especially in the case of the automatic ticketing device at the station described above. It is effective in an application that uses the recognition system that changes. However, this immediate speaker adaptation needs to be performed unsupervised using only a small amount of speech data.

In the conventional unsupervised speaker adaptation, an unspecified speaker is identified based on a decoding algorithm such as the Viterbi algorithm described in the document "Seiichi Nakagawa:" Speech Recognition by Stochastic Model ", IEICE, 1988". The input speech is recognized once using a reference model dictionary for a speaker (for example, a set of reference phoneme HMMs), and the phoneme HM of the input speech is recognized.
Estimate the M sequence. The phoneme HMMs are selected from the dictionary according to the phoneme HMM sequence 入 力 of the input speech estimated from the reference model dictionary and connected, and all the phoneme HMMs of the reference model dictionary are maximized so that the likelihood of the connected phoneme HMM sequence is maximized. The transformation parameter η in the function (model transformation function) Gη (θ) that maps the model parameter θ to the phoneme model parameter of the speaker to be recognized is estimated, for example, by maximizing the posterior probability (maxim
um a posteriori: MAP, for example JLGauvain and C.-H.Le
e, Maximum a posteriori estimation for multivariat
e Gaussian mixture observations of Markov chains,
IEEE Trans. Speech and Audio processing, Vol. 2, N
o.2, pp291-298, 1994).

[0011]

(Equation 1)

Here, X is an input voice, f () is a likelihood function, g
() Represents a prior probability density function. The parameter θ ′ of the phoneme HMM adapted to the speaker to be recognized is calculated using

[0013]

(Equation 2)

Is calculated as follows: Practically, for example, among the parameters θ representing the respective phoneme models,
Assuming that the variance σ, the weight coefficient w, and the state transition probability a do not change, only the average value m is adapted. Then, the model conversion function Gη () is

[0015]

(Equation 3)

[0016] In this way, the input speech X is re-recognized using the reference model dictionary adapted to the speaker to be recognized, and the recognition result is output. However, in the mapping of the model parameters θ using the equations (1) and (2), the phoneme HMM sequence Λ is correctly estimated for the input speech X of a speaker having low performance with respect to the unspecified speaker reference phoneme HMM dictionary. No, the effect of speaker adaptation was not always obtained.

[0017]

SUMMARY OF THE INVENTION It is a first object of the present invention to provide a speaker adaptation method for a speech model which can more accurately estimate a speech model sequence for input speech by speaker adaptation. It is. A second object of the present invention is to provide a speech recognition method having a high recognition rate using the above speaker adaptation method.

A third object of the present invention is to provide a recording medium on which the above-mentioned speech recognition method is recorded.

[0019]

According to the present invention, in a feature extraction process, an input voice of a speaker to be recognized is analyzed to extract a feature parameter sequence, and the extracted feature parameter sequence and an unspecified speech are extracted. In the model sequence extraction process, a plurality of hypotheses of the model sequence estimated to correspond to the feature parameter sequence of the input speech are extracted using the speaker's voice model, and the plurality of model sequences extracted in the temporary adaptation process are extracted. For each sequence, the connected speaker-independent speech model is adapted such that the likelihood between the model in which the speaker-independent speech models are linked according to the sequence and the feature parameter sequence of the input speech is maximized. In the adaptation model selection process, the likelihood between the model obtained by concatenating the adapted speech model for each of the model sequences according to the model sequence and the feature parameter sequence of the input speech is obtained. This by selecting the adapted speech models corresponding to the adaptation speech model to model sequence based on the likelihood.

In the adaptation model selection process, the adaptation speech is not based on the likelihood when using the speaker-independent speech model, but based on the likelihood using the speech model after speaker adaptation. We are selecting models. This is based on the principle that "even if the likelihood of the unspecified speaker's speech model for the correct model series is low, the likelihood of the speech model after speaker adaptation for that series is high." Is based on In the present invention, a plurality of model sequences are extracted so that a correct model sequence is included with a high probability, speaker adaptation is performed on each of these model sequences, and a speech model after each speaker adaptation is used. By comparing the likelihood with the recognition target speech at the time of occurrence and selecting the speech model having the highest value, the adaptation based on a more correct model sequence is performed.

In order to extract a plurality of model sequences so as to include a correct model sequence in the process of extracting a model sequence, for example, refer to the document “Automatic speech and spe, supervised by C.-H. Lee et al.
akerrecognition (Chapter 18, Multiple-pass search stra
tegies) ", Mu of Kluwer Academic Publishers, 1995"
N-best paradigm of ltiple-pass search strategies (pa
radigm) is available. Thereby, the search space of the model sequence can be effectively reduced. Specifically, using the speech of the speaker to be recognized, the parameters of the speech model for the unspecified speaker are adapted while performing the adaptation so as to increase the likelihood for the speech of the recognition target speaker. The speech model is adapted to the speaker to be recognized by re-selecting the hypothesis (model series) to be used from among the N-best hypotheses.

In the present invention, the adaptation to each model sequence is first performed roughly by strengthening the connection state between the parameters, and the roughly adapted speech model and the speech of the speaker to be recognized are converted. Adapts the speech model by performing adaptation for each of the multiple model sequences and selecting the one with the highest likelihood for the speech of the recognition target speaker when using the adapted speech model. I do. At this time, the adaptation to each model series is performed finely by loosening the connection between parameters compared to the previous time. Instead of adapting the speech model with the model parameters of one model sequence having the maximum likelihood selected in the adaptation model selection process, N
The speech model may be adapted with model parameters obtained by averaging the model parameters of the model series with weights corresponding to their likelihoods. At least one of the above procedures
Repeat several times. This repetition may be performed only in the temporary adaptation process and the adaptation model selection process. That is, a plurality of model sequences extracted first may be reused without returning to the model sequence extraction process.

When outputting the speech recognition result, the model sequence having the maximum likelihood finally selected in the adaptation model selection process is output as the recognition result. Alternatively, the input speech is re-recognized using the speech model after the adaptation corresponding to the model sequence selected in the adaptation model selecting process, and the model sequence showing the maximum likelihood is output as a recognition result. The recognition method using the adaptation algorithm according to the present invention is recorded in a recording medium in advance, and the recording medium can be used for various speaker-independent speaker recognition systems.

[0024]

BEST MODE FOR CARRYING OUT THE INVENTION

First Embodiment A first embodiment of the present invention will be described with reference to the flowchart of FIG. 2 and the functional block diagram of the speech recognition system of FIG.
The voice model storage unit 12 stores in advance a voice model for an unspecified speaker, for example, a voice HMM for an unspecified speaker, corresponding to a recognition category such as a word learned using the voices of many speakers as a reference model dictionary. It is assumed that Also,
The procedure for implementing the speech recognition method using the speaker adaptation method of the present invention described below is performed by the storage unit 10M in the control unit 10.
The control unit 10 controls the processes of the respective units 11 to 16 of the voice recognition system in FIG. 3 according to the procedure.

Step S1 (feature extraction process): The feature extraction unit 11 extracts the features of the voice data of the speaker to be recognized. In the feature extraction, the input audio data is subjected to LPC analysis for each fixed frame, and a time series of feature parameter vectors such as cepstrum or Δcepstrum is obtained as a feature parameter sequence. Step S2 (model sequence extraction process): The model sequence selection unit 13 connects the speech models selected from the model storage unit 12, and estimates that the model is closest to the speech data X converted into the feature parameter sequence obtained in step S1. Done,
Extract _N model sequences Λ ₁ , Λ ₂ ,... Λ _N. The extraction of this model series is described, for example, in the document “W. Chou et al .:“ An algorit
hm of high resolution and efficient multiple strin
g hypothesization for continuous speech recognitio
n using inter-word models ”, Proc.ICASSP, pp.II-15
3-156, 1994 ”.

More specifically, for example, the likelihood calculation unit 14 calculates the likelihood f (X | Λ, θ) of the assumed speech HMM sequence Λ from the speech data X converted into the feature parameter string X by the likelihood calculation unit 14. N are calculated in order from the one with the highest likelihood (N is 2
The model sequence Λ _n (n = 1,..., N) of the above-mentioned predetermined integer is selected and extracted from the model sequence selection unit 13. The number N is selected to such an extent that a model sequence (correct answer) that correctly represents the speech to be recognized is included with a high probability in the extracted N model sequences. For example, if N = 10 in the case of a 4-digit numeric voice (word sequence), the probability that the correct model sequence is included in these 10 model sequences is about 97%.
Choose 10 If the number of N is larger than this, the probability that a correct model sequence is included will be higher, and N should be larger if the amount of calculation and processing time are not considered. If the speech to be recognized becomes complicated, the number N needs to be increased accordingly. In the present invention, it is important that the extracted N model sequences have a high probability that a correct answer is included.

Step S3 (temporary adaptation process): Step S3 is performed on the speech data X converted into the characteristic parameter sequence.
For each model sequence Λ _n (n = 1,..., N) extracted in step 2,

[0028]

(Equation 4)

Thus, a conversion parameter η ′ _n that maximizes the likelihood function value f (X | Λ _n , η _n , θ) of the model sequence Λ _n with respect to the voice data X is obtained, and the conversion parameter η ′ _n is used. Speech HMM for unspecified speakers forming the model sequence Λ _n
Is converted to obtain θ ′ _n (= Gη ′ _n (θ)). As the method, for example, the Baum-Welch algorithm (for example, the document “Seiichi Nakagawa:“ Speech Recognition by Stochastic Model ”, IEICE, 1988)” or the MAP estimation algorithm can be used. With the model parameter θ ′ _n converted using this conversion parameter η ′ _n ,
For each of the N model sequences Λ _n (n = 1,..., N), all the speech HMMs constituting the model sequence are provisionally adapted to the speaker to be recognized.

[0030] That is, as shown in FIG. 4, model sequence lambda ₁
Is _composed of a sequence of speech models λ ₁₁ , λ ₁₂ ,..., Λ _1k1 , and the model sequence Λ ₂ is composed of speech models λ ₂₁ , λ _{22 ,.}
It is assumed that the model sequence Λ _N is composed of speech models λ _N1 , λ _N2 ,..., Λ _NkN . Likelihood function between the model sequence Λ ₁ and the voice data X when the conversion parameter η ₁ is changed according to equation (4)
A conversion parameter η ′ ₁ that maximizes f (X | Λ ₁ , η ₁ , θ ₁ ) is obtained, and using this η ′ ₁ , a model parameter θ of the unspecified speaker speech HMM is set to θ ₁ (= Gη ₁ (θ)). Note that all the speech models (HMM) λ constituting the model sequence Λ _{n
When the} conversion parameter η _n is determined in common for all the model parameters θ of _{n 1} , λ _n2 ,..., λ _{n kn} , the conversion of the parameters θ of the respective voice models is in a state in which they are mutually constrained. It is said that the state of knotting is strong.

Similarly, the speech model λ ₂₁ of the model sequence Λ ₂ ,
lambda _22, ..., lambda _2k2 changing the conversion parameter eta ₂ of the model series lambda ₂ of the likelihood function f (X for the audio data X | Λ _{2, η 2,}
theta ₂₎ is "seek _2, the eta 'transformation parameter eta becomes maximum ₂
Model parameters θ ₂ (= Gη ₂ (θ)) provisionally adapted from
Get. Similarly, the likelihood function for each model series Λ _n
A conversion parameter η ′ _n that maximizes f (X | Λ _n , η _n , θ _n ) is obtained, and a temporarily adapted model parameter θ _n is obtained. Thereby, N temporarily adapted speech model parameters θ ₁ ,
.., Θ _N are obtained.

Step S4 (adaptation model selection process):
Next, each model parameter provisionally adapted in step S3
TA θ_n Series Λ '_nFor each input speech data X,
In other words, the likelihood function f (X | Λ_n, θ_n), And these likelihood functions f
(X | Λ₁, θ₁),…, F (X | Λ_N, θ _NThe largest model in)
Le series Λ_iAs the correct model series. Follow
And the model series Λ_iIs the speaker
The voice HMM is adapted to the following.

As indicated by the dashed line in FIG. 2, the provisional adaptation in step S3 may be repeatedly performed continuously a predetermined number of times. In such a case, the provisional adaptation using the conversion parameter η ′ _n of each model series is finely performed for each iteration in a looser connection state (described later) than the previous provisional adaptation. That is, each model series Λ
_For every _n , all the model parameters θ of all the speech models that constitute it are classified into, for example, two groups (clustering), and converted so as to maximize the likelihood for the model parameters of the two groups (clusters). By controlling the parameters η _a and η _b independently, the equation
(2) the making of a provisional adaptability to model sequence lambda _n according (3). Practically, for example, the change amount Δm (η = Δm) of the average value m of the Gaussian distribution is used as the conversion parameter η as described above. In that case, the average value of all distributions included in all models constituting each model sequence lambda _n and clustered into two groups, the two change amounts of the average value in common for each cluster Delta] m _a, decide Delta] m _b . In this way, provisional adaptation is performed so that the likelihood is maximized for each of the N model sequences.

When outputting a speech recognition result, the model sequence Λ _i of the maximum likelihood selected in step S4 is output as a recognition result in step S5. Alternatively, as shown by the dashed line, the entire speech model of the reference model dictionary is adapted in step S6 by the conversion parameter η ′ _i finally used in the adaptation model sequence Λ _i selected in step S4,
Next, in step S7, the input speech data X is re-recognized by using the adapted reference model dictionary, one model sequence having the maximum likelihood is extracted, and the model sequence extracted in step S5 is recognized as a recognition result. Output as In this case, as the re-recognition processing, for the input speech data X converted into the feature parameter sequence, for example, by using the Viterbi algorithm, a model sequence in which the likelihood function f (X | Λ, θ) is maximized is determined. Output as Second Embodiment FIG. 5 shows a processing procedure of a second embodiment of the model adaptation method according to the present invention. In the second embodiment, the processing of the first steps S1 to S4 is the same as that of the first embodiment shown in FIG.
As in steps S1 to S4, the input voice is LP
C analysis and conversion into a feature parameter sequence, and in step S2, N speech model sequences having the highest likelihood for the input speech feature parameter sequence are selected from the reference model dictionary, and in step S3, each of the N model sequences is Adaptation is performed so that the likelihood becomes maximum, and a model sequence having the maximum likelihood is selected from the N provisionally adapted model sequences in step S4. After the end of step S4, the following processing is further executed.

Step S5 (N model adaptation process):
Model series selected in step S4 _iStep against
The transformation parameter η ′ giving the maximum likelihood in S3_iAre used in common.
Therefore, N models before provisional adaptation in the immediately preceding step S3
Adapt the sequence. Returning to step S3,
N model sequences adapted in step S5₁,…, Λ
_NThe likelihood of the input speech data X is again
Each model series to maximize₁,…, Λ_NTo
Tentative adaptation. Where the conversion parameter η_n By
Tentative adaptation results within each model series from the previous tentative adaptation.
And loosen the condition.

In step S4, a model sequence to which the maximum likelihood is given again from the N model sequences is selected as a correct model sequence. In the above-described second embodiment, step S
5, S3, and S4 may be further repeated a desired number of times, and the model sequence that has given the maximum likelihood in the last step S4 may be selected as a correct model sequence.

In the second embodiment, to output a speech recognition result, the model sequence finally selected in step S4 is output from the recognition result output unit 16 as a recognition result in step S6. Alternatively, as indicated by the dashed line, step S4
Step S immediately before the model series selected in
In step S7, all the speech models in the reference model dictionary are adapted using the same conversion parameter η ′ _i as given in step 3, and in step S8, the input speech data X is converted using the adapted reference model dictionary. The model is re-recognized, and one model sequence having the maximum likelihood is extracted, and is output as a recognition result in step S5. In this case, as the re-recognition processing, the input speech data X converted into the time series of the feature parameters is obtained, for example, by a Viterbi algorithm, to obtain a model series in which the likelihood function f (X | θ) is maximized. Output. Further, as in the case of the first embodiment in FIG. 2, the provisional adaptation in step S3 may be repeated a plurality of times, and then the process may proceed to the adaptation model selection process in step S4. Third Embodiment FIG. 6 shows a third embodiment of the model adaptation method according to the present invention.
4 shows a processing procedure of the embodiment. In the second embodiment described above, the model sequence selected in step S4 is added to the immediately preceding step S4.
The case where the N model sequences before provisional adaptation are adapted in step S5 using the same transformation parameter η′i as the transformation parameter η′i given in step 3 has been described. In the third embodiment,
In step S5, all models of the reference model dictionary are adapted, and the process returns to step S2 instead of step S3. Using the dictionary adapted in step S5, the N model sequences having the highest likelihood for the input speech data X are used. Perform the selection again.

That is, as in the case of the second embodiment shown in FIG.
A model sequence having the maximum likelihood in step S4 is selected from the provisionally adapted N model sequences as a provisionally correct model sequence. In step S5 in the third embodiment, and adapting the entire model of the reference model dictionary with the 'same transformation parameter η and _{i' i} transformation parameter η given in step S3 to the selected model sequence returns to step S2. Using the reference model dictionary adapted in step S2, the N model sequences having the highest likelihood with respect to the input speech data X are again selected, and in steps S3 and S4, provisional adaptation and adaptation are performed as in the previous case. Select a model.

Further, steps S5, S2, S3 and S4 may be repeated. Also in this case, it is preferable to perform the hierarchical adaptation of the model sequence in step S3 by performing the hierarchical clustering of the model parameters for each repetition process as described above. In this case, in the adaptation of the reference model dictionary in step S5, the same clustering as the clustering of the model parameters in step S3 is performed on the parameters of all the models in the dictionary, and the model parameters θ
Are grouped, and conversion parameters for each cluster determined in step S3, such as η _ia and η _ib
Is independently applied to the model parameters of each group to adapt the dictionary.

In the repetition processing of steps S5, S2, S3, and S4, the model parameters of each cluster are further divided into, for example, two groups by clustering the model parameters for each model series in the previous step S3 in step S3 for each repetition. Perform clustering in a hierarchical manner like clustering. Therefore, in the third model adaptation, the transformation parameters η _ia , η _ib ,
Adaptation is performed by giving η _ic and η _id . Similarly, in the clustering of model parameters for all models in the dictionary in step S5, the number of clusters is hierarchically increased each time the processing loop is repeated, thereby loosening the connected state of the parameters between the models. In step S4 in the last round, the model series with the highest likelihood is selected as the correct model series from the respective model series after provisional adaptation obtained in the immediately preceding step S3. Further, as indicated by a line in FIG. 6, the provisional adaptation in step S3 may be repeated for a predetermined number of consecutive times in which the connection between the conversion parameters is loosened compared to the previous time.

When outputting the recognition result, the model sequence having the maximum likelihood selected in step S4 in the last repetition is output. Alternatively, after the repetitive adaptation process of steps S2 to S5 is performed a predetermined number of times, the process ends in the reference model dictionary adaptation process of step S5, and then, as indicated by the dashed line, the adapted reference model dictionary in step S6 May be used to re-recognize the input voice data X, extract one model sequence having the maximum likelihood, and output it as a recognition result. In this case, as the re-recognition processing, the input voice data X converted into the time series
The bi-algorithm obtains a model sequence with the maximum likelihood function f (X | θ) and outputs it as a recognition result. Also,
As in the case of the first embodiment shown in FIG. 2, the provisional adaptation in step S3 may be repeated a plurality of times, and then the process may proceed to the adaptation model selection process in step S4.

In the first, second, and third embodiments, the step
In step S3, f (X | Λ_n, η _n, θ)
F (X | Λ)_n, η_n, θ) g
(θ) may be used. In the above description,
Is not limited to HMM. Also, the model system in step S2
Extract columns only when they follow the N-best paradigm
The point is that the correct model series is included with high probability
You only need to extract it.

In the above-described second and third embodiments, step S3
In, temporarily adapting the selected N model sequence lambda ₁ to [lambda] _N according N-best method, among them at step S4, selects one model sequence which showed the greatest likelihood, and the selected model Although other models have been adapted using the transformation parameters of the series, the correct model series does not always show the maximum likelihood due to the temporary adaptation. H for unspecified speakers
When the MM is used, especially when the correct model series is recognized lower than 3-best, the correct answer often does not become the first place even if provisional adaptation is performed. Therefore, as shown below, instead of selecting only one model sequence having the maximum likelihood from the N model sequences in step S4, the lower model sequence is considered to contribute to the adaptation in consideration of the reliability. It may be.

[0044] That is, in this modification, in step S4, the reliability _{C n (n = 1, ...} , N) for the highest likelihood of N model sequence lambda ₁ to [lambda] _N selected in step S2 the following Define by expression. C _n = {f (X | Λ _n , θ _n ) / f _max } ^r where f _max is the maximum likelihood of the input speech data X indicated by the N model sequences Λ _n after the tentative adaptation. _{things, f (X | Λ n,} θ n) is the likelihood indicated by each model sequence lambda _n after temporary adaptation, r is a constant parameter set experimentally. This reliability C _n is calculated for each series, and the speech HM
The model parameters of M are calculated according to the following equation.

[0045] performing the adaptation of theta step S5 with a _{^{'= Σ N n = 1 C}} n θ n / Σ N n = 1 C n parameter theta obtained in this way'. This method is called a smoothing estimation method.

[0046]

As described above, in the present invention, a plurality of model sequences of an unspecified speaker's speech model are extracted using the speech of the speaker to be recognized, The voice model for a specific speaker is adapted. Therefore, if the correct answer is included in the extracted model series,
It is highly probable that the likelihood between the model sequence adapted for each model sequence in step S3 and the speech data X will be the largest for the correct model sequence. In the present invention, since the adapted speech model based on the maximum model sequence is adopted, more correct adaptation is realized. That is, for example, if the model sequences Λ ₁ ,..., Λ _N are extracted in order from the one with the highest likelihood in step S2, if the likelihood of the correct model sequence is lower than the first, Adaptation of a speech model is performed using a model sequence that is not. However, in the present invention, steps S3 and S4 increase the possibility that the likelihood with the adapted speech model based on the correct model sequence is maximized, and more correct adaptation is obtained.
Therefore, as described above, it is important that the probability that a correct answer is included in the N model sequences extracted in step S2 is high. The effect of the present invention was examined in a four-digit number (word series) recognition experiment. This experimental example will be described. The speech HMM used in the experiment is a continuous HMM with Gaussian mixture distribution number 4.
It is. 13 types of voice HMMs (/ rei /, / maru /, / zero /, / ichi /, / ni /, / san /, / y)
on /, / go /, / roku /, / nana /, / hachi /, / kyu /, / ku /)
Was prepared. To create a voice HMM for unspecified speakers, a man 17
HMM parameters were estimated by Baum-Welch algorithm using a total of 24,194 utterances (isolated numbers, two-digit numbers, or four-digit numbers) by seven persons. For speaker adaptation and recognition, six utterances (4
(Digit number) was evaluated by the 4-digit number recognition rate. As feature parameters, audio data having a sampling frequency of 8 kHz is converted to a frame length of 32 m for every 8 ms of frame period.
A cepstrum and a Δcepstrum were extracted by performing LPC analysis of order 12 with s.

In the speaker adaptation, only the bias component Δm (ie, η = Δm) of the average value m of each Gaussian mixture distribution of the speech HMM is estimated in accordance with the third embodiment, and the total number of bias Δm is initially set for all models. 1 for all distributions (strong knot state)
Each time the adaptation process was repeated, it was hierarchically estimated as 2, 4, 8, and 16 (that is, clustering was performed hierarchically). The number N of the N-best hypotheses (the extracted model sequence in step S2) was set to 10.

FIG. 7 shows an unspecified speaker according to the recognition rate (baseline) using the unspecified speaker's speech HMM and a model sequence obtained by performing speech recognition using the unspecified speaker's speech HMM. The recognition rate using the speech HMM adapted by the conventional adaptation method for speaker adaptation of the speech HMM for use (conventional method) and the speech HMM adapted by the method of the third embodiment of the present invention are used. And the recognition rate (this method). From these results, it can be seen that the method of the present invention is more effective than the conventional method.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining an HMM, in which A is a characteristic parameter sequence indicating three characteristic regions of a phoneme, B is a schematic mixed Gaussian distribution of characteristic parameters in each region, C
Indicates a state transition between regions.

FIG. 2 is a flowchart showing a procedure of a method according to a first embodiment of the present invention.

FIG. 3 is a functional block diagram of a speech recognition system embodying the present invention.

FIG. 4 is a diagram showing an extracted model sequence obtained in step S2 and a corresponding likelihood function.

FIG. 5 is a flowchart showing a procedure of a method according to a second embodiment of the present invention.

FIG. 6 is a flowchart showing a procedure of a method according to a third embodiment of the present invention.

FIG. 7 is a view showing an experimental result showing an effect of the present invention.

Claims (14)

[Claims] 1. A reference model dictionary for an unspecified speaker which learns using the voices of a large number of speakers and constructs a speech model corresponding to a recognition category such as a phoneme or a word using model parameters. In the speaker adaptation method of a speech model that adapts a speech model to the speech of a speaker to be recognized, (a) feature amount extraction for extracting a speech feature parameter sequence of the input speech of the speaker to be recognized (B) extracting, from the reference model dictionary, a plurality of hypothesis model sequences estimated to correspond to the feature parameter sequence of the input speech from the unspecified speaker voice model, c) For each model sequence of the plurality of extracted hypothesis model sequences, the model of each of the above-mentioned hypotheses is maximized so that the likelihood of the model sequence of the hypothesis with respect to the feature parameter sequence of the input speech is maximized. A tentative adaptation process of tentatively adapting by controlling the model parameters of the Dell sequence, and (d) based on the likelihood of the respective hypothesis model sequences after the tentative adaptation in the tentative adaptation process with respect to the feature parameter sequence. And a step of selecting at least one of the hypothetical model sequences after the provisional adaptation as an adapted speech model sequence. 2. The speaker model adaptation method according to claim 1, wherein said adaptation model selecting step (d) is performed after said hypothesis model sequence is provisionally adapted in said provisional adaptation step (c). And selecting a hypothetical model sequence having the maximum likelihood for the feature parameter sequence as the adapted speech model sequence. 3. The speaker adaptation method for a speech model according to claim 1 or 2, further comprising the step of selecting said unspecified speech based on model parameters of an adapted speech model sequence selected in said adaptation model selecting step (d). Adapting the user's speech model sequence, and performing the tentative adaptation process (c) by, for each of the plurality of extracted hypothesis model sequences, the likelihood of the sequence with respect to the extracted feature parameter sequence. Using the adapted speaker-independent speech model adapted to make the connection between the conversion parameters looser than the previous time, and performing the finer adaptation, and selecting the adaptation model And re-selecting the adapted model sequence in the step (d) again at least once. 4. The speaker adaptation method for a speech model according to claim 1, wherein the speech model of the reference model dictionary is added to the temporary adaptation model sequence selected in the adaptation model selection step (d). A reference model dictionary adaptation process to be adapted, and using the adapted reference model dictionary, in the model sequence extraction process, again extract the plurality of hypothesis model sequences having the maximum likelihood for the input speech, A step of loosely tying the connection state between the parameters in the provisional adaptation process (c) and provisionally adapting the hypothesis model sequence in detail, and among those provisionally adapted hypothesis model sequences,
The method further includes the step of selecting a readapted model sequence having the maximum likelihood. 5. The speaker adaptation method for a speech model according to claim 2 or 4, wherein the provisional adaptation step (c) is performed by fine adaptation by loosening the connection state between the conversion parameters as compared with the previous time. The method further includes a step of continuously repeating the above operation a predetermined number of times. 6. The speaker adaptation method for a speech model according to claim 1 or 2, wherein the model sequence extraction step (b) includes the step of extracting a plurality of model sequences having the greatest likelihood with the feature parameter sequence of the input speech. It is the process of getting. 7. The speaker adaptation method for a speech model according to claim 1, wherein the adaptation model selecting step (d) is provisionally adapted in the provisional adaptation step for the feature parameter sequence of the input speech. And selecting the speech model as a sum of the adapted speech models corresponding to the model sequence by weighting corresponding to the likelihood of the extracted model sequence. 8. The method according to claim 1, wherein the speech model is a speech HMM defined by a statistic parameter. 9. The speaker adaptation method for a speech model according to claim 8, wherein the adaptation or the provisional adaptation controls an average value of a Gaussian distribution which is at least one of model parameters defining the speech HMM. This is a process of adapting so that the likelihood is maximized. 10. A speaker model adaptation method according to claim 1 or 2, wherein a recognition category corresponding to the model sequence selected in said adaptation model selection step is determined with respect to said input speech of said recognition target speaker. A speech recognition method including a step of outputting as a speech recognition result. 11. The speaker adaptation method for a speech model according to claim 2, wherein all the speech models in the reference model dictionary are adapted to a reference model dictionary corresponding to a model sequence selected in the adaptation model selection process. Extracting the speech model sequence having the highest likelihood for the feature parameter sequence of the input speech from the step and the adapted reference model dictionary,
A re-recognition step of outputting as a recognition result. 12. The method according to claim 4, wherein the characteristic parameter sequence of the input speech is re-used by using the last adapted speech model for the unspecified speaker in the reference model dictionary. A speech recognition method including a step of performing a recognition process, extracting one model sequence that gives the maximum likelihood, and outputting the model sequence as a recognition result for the input speech of the speaker to be recognized. 13. The speaker adaptation method for a speech model according to claim 11, wherein the re-recognition processing includes converting the input speech data converted into a time series of feature parameters into a Viterb speech data.
This is a process for obtaining a model sequence with the maximum likelihood function by the i-algorithm. 14. A recording medium in which the procedure of the voice recognition method according to claim 10, 12 or 13 is recorded.