JPH0627989A - Background hmm parameter extracting method and speech recognizing device using the same - Google Patents

Abstract

PURPOSE:To learn a background HMM(Hidden Markov Model) matching a purpose at a high speed and speedily converge a parameter. CONSTITUTION:A series of labels for learning specific speeches used to learn the background HMM is supplied from a speech quantization part 2 to an HMM parameter estimation part 91, which estimates the parameters of the HMM, at every speech and stores them in an HMM parameter storage part 92. When the parameters of the background HMM are found, the parameters of the HMM at every speech used to learn the background HMM are taken out of the HMM parameter storage part 92, a background HMM parameter arithmetic part 93 calculates, for example, the mean of the respective parameters, and the result is stored as the parameter of the background HMM in a background parameter storage part 6.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a background HMM suitable for extracting parameters of a background HMM used in combination with a keyword HMM for recognition of input speech.
The present invention relates to an MM parameter extraction method and a speech recognition device using the method.

[0002]

2. Description of the Related Art In recent years, as a speech recognition method for recognizing input speech, vector quantization or matrix quantization for converting speech into a fixed code sequence is performed, and the quantized code sequence is hidden Markov model. , Ie HMM
The method of recognizing with is successful.

In order to make a voice recognition device to which this kind of voice recognition system is applied more practical, an input voice in which unrecognized words and ambient noise are added before and after the word to be recognized. On the other hand, word spotting technology that correctly recognizes a word to be recognized is essential.

In order to perform word spotting, it is necessary to compare HMM scores for a large number of different voice section candidates. However, since the voice sections to be compared are different, the score varies greatly depending on the voice section, and high recognition performance cannot be obtained if the score obtained from the HMM of each recognition target word is directly used for the recognition processing.

Therefore, as a method of normalizing the score of the HMM suitable for performing the word spotting, the speech (word) to be recognized, that is, the logarithmic likelihood Lk of the keyword HMM (hereinafter referred to as the keyword HMM) is recognized. The difference from the log-likelihood Lb of the HMM (hereinafter, referred to as background HMM) of the target speech (word) or noise is calculated,
How to use this difference as the score of HMM is Richard of MIT in the US
d Reported by C. Rose, etc. (ICASSP-90, S2.2
Four).

Conventionally, the parameters of this background HMM have been determined (estimated) by inputting and learning various different voices and noises other than keywords into one HMM.

[0007]

As described above, the background HMM used in the conventional method for normalizing the score of the HMM suitable for performing the word spotting is conventionally used.
The parameters have been obtained by inputting various different voices and noises other than the keywords into one HMM and learning them.

For this reason, when the type of voice used for learning the background HMM is changed, the parameters of the background HMM must be learned again, which is a problem. Further, since different voices are input to one HMM to obtain the parameters, there is a problem that the parameters do not converge.

Therefore, the present invention provides a background HMM.
For each speech or noise not used for recognition used for learning, the parameters of the HMM peculiar to the speech or noise are found and stored, and the parameters of the background HMM are calculated from the parameters found for each speech or noise. , The problem of poor convergence of parameters can be solved, and the background HMM
Even when changing the type of speech or noise used for learning, it is only necessary to find the parameters of the HMM specific to the speech or noise to be changed, and the background that suits the purpose can be learned quickly. H
An object of the present invention is to provide an MM parameter extraction method and a speech recognition device using the method.

[0010]

The present invention relates to each of various arbitrary voices or noises used for learning (parameter extraction) of a background HMM used in combination with a keyword HMM for recognizing an input voice. Alternatively, HMM parameters specific to noise are obtained and stored in the storage means, and each H obtained for each voice or noise is obtained.
It is characterized in that the parameters of the background HMM used for obtaining the log-likelihood for the background HMM are obtained based on the parameters of the MM.

Further, according to the present invention, when various voices or noises used for learning the background HMM are changed, the HM is stored in the storage means among the various voices or noises after the change.
Only for speech or noise for which the M parameter is not stored, the HMM parameter specific to the speech or noise is obtained and stored in the storage means, and among the parameters of each HMM stored in this storage means It is also characterized in that the parameters of the background HMM are re-extracted based on the parameters of each HMM of various voices or noises. The present invention is also characterized in that the above-mentioned background HMM parameter extraction function is incorporated in a voice recognition device.

[0012]

In the above structure, the background HM
The parameters of the HMM required for learning M are obtained for each voice (or noise) other than the keyword (the voice that is the recognition target), and from the HMM parameters obtained for each voice (or noise) Background HM
The parameters of M are determined. For the calculation at this time, a method of taking an average or a weighted average of the parameters of the HMM obtained for each voice (or noise) is applied.

As described above, in the above configuration, the parameters of the background HMM are obtained from the HMM obtained for each voice (or noise) required for learning the background HMM, so that the background HMM can be learned. When the voice (or noise) to be used is changed, only the HMM parameters of the voice (or noise) that have not yet been obtained are obtained, and the parameters of each HMM of each voice (or noise) after the change are obtained. It is possible to re-extract the parameters of the background HMM based on Therefore, it is not necessary to redo the calculation of the parameters of the background HMM for all the changed voices, and the background HMM can be learned at high speed.

Further, in the above configuration, each voice (or noise) used for learning the background HMM is not directly used for parameter calculation of the background HMM as in the conventional case, but each voice (or noise) is used. To H
After the parameters of the MM are calculated, the parameters of the background HMM are calculated using these parameters, so that the parameters converge quickly and learning can be performed at high speed.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a voice recognition apparatus according to an embodiment of the present invention. The processing in the speech recognition apparatus of FIG. 1 basically uses a phonetic segment (hereinafter referred to as PS) as a processing unit for recognition in terms of phonetically, and a PS is recognized by using the recognition dictionary of the PS unit. The HMM recognition unit performs word matching on the quantized input speech. The processing up to this point in the speech recognition apparatus of FIG. 1 will be described in more detail.

First, a voice signal (input voice) input through a microphone (not shown) is guided to the acoustic analysis unit 1. The acoustic analysis unit 1 is for acoustically analyzing the input voice to obtain characteristic parameters. The acoustic analysis unit 1, as shown in detail in FIG. 2, includes an A / D converter 11, a power calculation unit 12, and an LPC analysis unit 13.

The input voice guided to the acoustic analysis unit 1 is A / D
In the converter 11, for example, a sampling frequency of 12 kHz
After being quantized with z, 12 bits, it is input to the power calculation unit 12, the voice power thereof is calculated, and further input to the LPC analysis unit 13, and LPC (Linear Predictive Co
ding) Mel cepstrum analysis (LPC analysis). This LPC analysis is performed using a 16th-order LPC mel cepstrum as an analysis parameter with a frame length of 16 msec and a frame period of 8 msec, for example. The acoustic analysis by the acoustic analysis unit 11 is not limited to the LPC mel cepstrum analysis, and may be a BPF (Band Pass Filter) analysis or the like.

The characteristic parameters of the input voice obtained by the acoustic analysis in the acoustic analysis unit 1 are sent to the voice quantization unit 2. The voice quantizing unit 2 obtains a label (PS label) for each frame from the characteristic parameter by using the PS dictionary storage unit 3 in which a recognition dictionary of a predetermined PS (voice segment) unit is registered. That is, the voice quantization unit 2 uses the PS of the characteristic parameters of the input voice analyzed by the acoustic analysis unit 1.
A matching process with the recognition dictionary of a predetermined PS unit registered in the dictionary storage unit 3 is continuously performed in the time axis direction, and the PS having the maximum similarity for each frame is output as the quantization result. Here, the continuous matching process by PS in the voice quantization unit 3 is performed using the composite LPC mel-cepstral similarity measure shown in the following expression (1).

[0020]

[Equation 1]

In the equation (1), C is the LPC mel cepstrum, W _m ^(Ki) and φ _m ^(ki) are the PS names K, respectively.
These are the weights and eigenvectors found from the eigenvalues of i.
In addition, (・) indicates the inner product and ‖ ‖ indicates the norm. Examples of PS used in this embodiment include the following. (1) Persistence segment: (1-1) Vowel stationary part (1-2) Friction consonant part (2) Consonant segment: Part including transition (transition part) to vowel [semi-syllable] (3) Syllable boundary segment : (3-1) Vowel boundary (3-2) Vowel, consonant boundary (3-3) Vowel, silence boundary (4) Other segment: unvoiced vowel, etc.

Of these, some of (1), (2) and (4) are often adopted when syllables are used as recognition segments. However, the PS of this embodiment is characterized in that the syllable boundary segment of (3) above is adopted in addition to the segments of (1), (2) and (4) above.

Now, the PS label sequence obtained by the continuous matching process in the voice quantizer 2 is the HMM recognizer 4
Sent to. The HMM recognizing unit 4 uses the label sequence obtained by the voice quantizing unit 2 as a keyword HMM for each of a plurality of types of speech (keywords) to be recognized and a background HMM (common to each keyword HMM). To recognize. The parameters of each keyword HMM are obtained in advance by the keyword HMM parameter estimation unit 8 described later and stored in the keyword HMM parameter storage unit 5. On the other hand, the background HMM parameter common to each keyword HMM is obtained by the background HMM parameter extraction unit 9 described later and stored in the background HMM parameter storage unit 6. It is also possible to separately prepare the background HMM for each keyword HMM.

The HMM recognition unit 4 calculates the logarithmic likelihood Lk of each keyword HMM by the keyword HMM likelihood calculation unit 4.
1. Background HMM likelihood calculation unit 42 for calculating log likelihood Lb of background HMM, score normalization for normalizing scores for each keyword HMM based on these two types of log likelihood Lk, Lb It is configured by a unit 43 and a score comparison unit 44 that performs a comparison calculation of each score normalized by the score normalization unit 43.

Now, from the voice quantizer 2 to the HMM recognizer 4
The label sequence sent to the recognition unit 4 is the keyword H in the recognition unit 4.
It is passed to the MM likelihood calculator 41 and the background HMM likelihood calculator 42.

The keyword HMM likelihood calculation unit 41 calculates the log likelihood Lk for each keyword HMM of the label sequence obtained by the speech quantization unit 2 for each keyword HMM stored in the keyword HMM parameter storage unit 5. It is obtained using parameters (output probability, transition probability, etc., which will be described later).

On the other hand, the background HMM likelihood calculation unit 42 stores the log likelihood Lb for the background HMM of the label sequence obtained by the voice quantization unit 2 in the background HMM parameter storage unit 6. It is obtained using the parameters of the ground HMM (output probability, transition probability, etc.).

The log-likelihoods Lk and Lb obtained for each of the keyword HMM and the background HMM by both the calculation units 41 and 42 are sent to the score normalization unit 43. The score normalization unit 43 calculates the log likelihood Lk of each keyword HMM.
And background H (common to each keyword HMM)
The difference between the log-likelihoods Lb of the MMs is obtained, and the score of each keyword HMM is normalized. The score of each keyword HMM normalized by the score normalization unit 43 is compared and calculated by the score comparison unit 44, and the comparison calculation result is passed to the recognition result output unit 7.

The recognition result output unit 7 outputs the voice recognition result for the input voice, and the score becomes maximum based on the result of the comparison operation in the score comparison unit 44 (corresponding to the keyword HMM). The keyword is output as the recognition result.

Here, a general formulation of the HMM will be described. It is assumed that the HMM has N states S ₁ , S ₂ , ..., _SN , and the initial state is stochastically distributed to these N states. For speech, a model that transitions a state with a certain probability (transition probability) is used for each fixed frame period.
At the time of transition, a label is output with a certain probability (output probability), but a null transition that transitions the state without outputting the label may be introduced. Even if the output label sequence is given, the state transition sequence is not uniquely determined. Since only label sequences can be observed, hidden markov mod
It is called el (HMM). The HMM model M is defined by the following six parameters. N: number of states (states S ₁ , S ₂ , ..., S _N ) K: number of labels (labels R = 1, 2, ..., K) p _ij : transition probability Probability of transition from S _i to S _j q _ij ( k): Probability of outputting the label k at the transition from S _i to S _j m _i : Probability of initial state S Probability that initial state is S _i F: Set of final states

Next, the model M is subject to transition restrictions that reflect the characteristics of voice. In voice, the state S is generally
The transition of the loop from _i to the previous passed state (S _i-1 , S _i-2 , ...) Is not allowed because it disturbs the temporal context. As a structure of this kind of HMM, an example as shown in FIG. 3 is typical.

The HMM is evaluated by the probability Pr that the model M outputs the first-ranked label sequence O ₁ = o ₁₁ , o ₂₁ , ..., O _T1.
(O / M). At the time of recognition, the HMM recognition unit 4 assumes each model and searches for a model M that maximizes Pr (O / M). Each model assumed in the HMM recognizing unit 4, that is, the model (parameter thereof) of the keyword HMM is obtained by learning the keyword HMM and is stored in the keyword HMM parameter storage unit 5.

The speech recognition apparatus of FIG. 1 is provided with the keyword HMM parameter estimation unit 8 so that the keyword HMM can be learned. This keyword HM
In order to learn the keyword HMM using the M parameter estimation unit 8, the speaker (voice) is sequentially uttered by the speaker while the speech recognition apparatus of FIG. 1 is set to the learning mode of the keyword HMM. Ask. The input voice from this speaker is received and LPC analysis is performed by the acoustic analysis unit 1 to obtain a characteristic parameter, and the PS quantization sequence is obtained from the characteristic parameter by the speech quantization unit 2 as learning data.

A plurality of speakers perform the above-mentioned keyword voice utterance on a plurality of keywords, so that the learning PS for each speaker for the same keyword can be obtained.
A label series is calculated for a plurality of predetermined keywords. These many learning PS label sequences are passed to the keyword HMM parameter estimation unit 8 via a storage unit (not shown).

The keyword HMM parameter estimation unit 8 is
The above learning PS label sequence is given to the HMM for each keyword, and the parameters of the model M corresponding to each keyword are estimated so as to maximize Pr (O / M). Then, the keyword HMM parameter estimation unit 8 registers the estimated model parameter for each keyword in the keyword HMM parameter storage unit 5. Next, the background HMM parameter extraction directly related to the present invention will be described.

First, the speech recognition apparatus of FIG. 1 is provided with a background HMM parameter extraction unit 9 so that the parameters of the background HMM can be extracted. This background HMM parameter extraction unit 9
Is an HM for learning the parameters of the HMM for each of a plurality of types of speech (or noise) (which are not recognized) necessary for extracting the parameters of the background HMM.
The parameter of the HMM obtained by the M parameter estimation unit 91 and the estimation unit 91 is set to the corresponding speech (or noise).
An HMM parameter storage unit 92 for storing each, and a background HMM parameter calculation unit that extracts the HMM parameter of the voice (or noise) to be the background from the storage unit 92 and calculates the parameter of the background HMM. 93 and a background model management unit 94. The management unit 94 controls the HMM parameter estimation unit 91 and the background HMM parameter calculation unit 93.

In this embodiment, the background HM
In order to obtain the parameters of the background HMM by the M parameter extraction unit 9, it is necessary to set the speech recognition apparatus of FIG. 1 to the learning mode of the background HMM. Then, in this state, the learning data about the voice necessary for the target parameter extraction is given to the background HMM parameter extraction unit 9 through the acoustic analysis unit 1 and the voice quantization unit 2.

Here, assuming that the voices necessary for learning the background HMM, that is, the voices not to be recognized as the background are three voices A, B, and C, the voices A, B, and C are given. A large number of learning data about
It may be given to the background HMM parameter extraction unit 9.

The HMM parameter estimation unit 91 in the background HMM parameter extraction unit 9 uses the above-mentioned speech A,
From a large number of learning data about B and C, speech A, B,
For each C, (similar to the parameter estimation in the keyword HMM parameter estimation unit 8), its voice A, B, C
Estimate the parameters of the HMM. Then, the HMM parameter estimation unit 91 stores the estimated HMM parameters of the voices A, B, and C in an area inside the HMM parameter storage unit 92 designated by the background model management unit 94 for each voice.

The background HMM parameter calculation unit 93 in the background HMM parameter extraction unit 9
Is a voice A necessary for learning the background HMM under the designation of the background model management unit 94.
The HMM parameters of B and C are retrieved from the HMM parameter storage unit 92. Then, the background HMM parameter calculation unit 93 calculates, for example, an average of the extracted parameters of the voices A, B, and C to calculate the background HMM.
The MM parameters are calculated and stored in the background HMM parameter storage unit 6.

As described above, in this embodiment, first, the HMM parameters are obtained for each voice necessary for learning the background HMM and stored in the HMM parameter storage unit 92, and thereafter, the background HMM learning is performed. The background HMM parameters are obtained by calculation based on the HMM parameters of each voice required for the above.

Therefore, the voices used for learning the background HMM are changed from the voices A, B, and C to, for example, the voice A,
When changing to B and D, it is not necessary to re-estimate the HMMs of the voices A and B (since the parameters are stored in the HMM parameter storage unit 92), and a new voice D is obtained.
It is only necessary to estimate the parameters of the HMM. This will be specifically described.

First, the HMM parameter estimation of the voice D is performed by giving the learning data of the voice D to the HMM parameter estimation unit 91, as in the case of the estimation of the HMM parameters of the voices A, B and C described above. Be seen. The HMM parameters of the voice D estimated by the HMM parameter estimation unit 91 are stored in the internal area of the HMM parameter storage unit 92 specific to the voice D designated by the background model management unit 94.

The background HMM parameter calculation unit 93, under the designation of the background model management unit 94, sets the parameters of the respective HMMs of the voices A and B, which have already been obtained from the HMM parameter storage unit 92, and this time. The parameters of the HMM of the voice D obtained in step 1 are taken out, the average of them is calculated, and the parameter of the background HMM after the change is obtained. Then, the background HMM parameter calculation unit 93 rewrites the contents of the background HMM parameter storage unit 6 with the newly obtained parameters of the background HMM.

As in the conventional case, the background H
In the method of directly estimating the parameters of the background HMM by inputting all the voices used for MM learning into one HMM, in the above example, the estimation of the parameters is performed by using all the voices A, B, and D. This has to be done, which greatly increases the amount of calculation. Further, when data of different voices are input to one model, the convergence of parameters becomes poor.

Now, the gist of the present invention is that when obtaining the parameters of the background HMM, which is a model of unrecognized speech, each different speech (or noise) used for learning the background HMM is treated as one. After estimating the parameters of the model for each speech (or noise), instead of inputting them into the model and estimating the parameters directly, the background HMM is calculated from each of these parameters.
To find the parameters of

Therefore, the structure of the HMM is not limited to that shown in FIG. In the above embodiment, the average of the HMM parameters for each voice is calculated as the background HMM.
However, the present invention is not limited to this. For example, the weighted average method is applied, the parameters of the HMM for each voice are weighted, and then the average is calculated to obtain the parameters of the background HMM. You may do it.

Further, the keyword HMM parameter estimation unit 8 and the background HMM parameter extraction unit 9
Is not necessarily provided in the voice recognition device, and the keyword HMM parameter estimation unit 8 and the background HMM parameter extraction unit 9 are used in another device to obtain the keyword HMM parameter and the background HMM parameter, It may be stored in the keyword HMM parameter storage unit 5 and the background HMM parameter storage unit 6 through a disk device or the like.

In the above embodiment, the unit for quantizing speech is PS (speech segment), but the unit of quantization may be a phoneme or a syllable, and a clustering method different from acoustic classification. You may use the unit quantized using. In addition, the present invention can be variously modified and implemented without departing from the scope of the invention.

[0050]

As described above, according to the present invention, the parameters of the background HMM are obtained from the parameters of the HMM stored in the storage means obtained for each voice (or noise) used for learning the background HMM. Since the calculation is performed, when the voice used for learning the background HMM is changed, it is not necessary to recalculate the parameters of the background HMM for all voices, and the background HMM suitable for the purpose can be obtained.
Can be learned at high speed.

According to the present invention, instead of directly using different voices for the background HMM parameter calculation, the HMM parameters are calculated for each voice, and the background HMM parameters are calculated using the calculation results. Since, the parameters converge quickly and learning can be performed at high speed.

[Brief description of drawings]

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

2 is a block diagram showing a detailed configuration of an acoustic analysis unit 1 in the apparatus shown in FIG.

FIG. 3 is a diagram showing a typical example of the structure of an HMM.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Acoustic analysis part, 2 ... Speech quantization part (label extraction means), 4 ... HMM recognition part, 5 ... Keyword HMM parameter storage part, 6 ... Background HMM parameter storage part, 7 ... Recognition result output part, 8 ... Keyword HMM parameter estimation unit, 9 ... Background HMM parameter extraction unit, 41 ... Keyword HMM likelihood calculation unit, 42 ... Background HMM likelihood calculation unit, 43 ... Score normalization unit, 44 ... Score comparison unit, 91 ... HMM Parameter estimation unit, 92 ... HMM parameter storage unit (storage unit), 93
… Background parameter calculator.

Claims (5)

[Claims] 1. A label is calculated for each frame from a feature parameter obtained by acoustically analyzing an input speech signal, and a log-likelihood of a sequence of this label with respect to an HMM (Hidden Markov Model) of a keyword to be recognized. And a log likelihood of a background HMM which is an HMM of any various speech or noise other than the keyword, of the same label series as the label series, and the keyword score based on these two log likelihoods. In the speech recognition method for generating a speech and recognizing an input speech, a HMM parameter specific to each speech or noise is obtained for each of various kinds of speech or noise used for parameter extraction of the background HMM. Based on the parameters of each HMM stored in the storage means and obtained for each voice or noise. , The background H
A background HMM parameter extraction method, characterized in that a parameter of the background HMM used for obtaining a log-likelihood for MM is obtained. 2. When changing various voices or noises used for the parameter extraction of the background HMM, only the voices or noises for which the parameter is not stored in the storage means among the various voices or noises after the change , HMM parameters specific to the voice or noise are obtained and stored in the storage means, and among the parameters of the HMMs stored in the storage means, the HMM parameters of various voices or noise after change are also stored. The background HMM according to claim 1, wherein the parameters of the background HMM are extracted again.
Parameter extraction method. 3. The background according to claim 1, wherein the parameter of the background HMM is obtained by taking an average or a weighted average of the parameters of the HMM for each of the various voices or noises. HMM parameter extraction method. 4. An acoustic analysis means for obtaining a characteristic parameter by acoustically analyzing an input voice signal, a label extraction means for obtaining a label for each frame from the characteristic parameter obtained by the acoustic analysis means, and this label extraction. Of the series of labels found by
First likelihood calculating means for obtaining a logarithmic likelihood of a keyword to be recognized with respect to an HMM, and a logarithm of a background HMM which is an HMM of any kind of speech or noise other than the keyword, of the same label series as the label series. Second likelihood calculating means for obtaining a likelihood, and score calculating means for obtaining a score of a keyword to be recognized based on both calculation results of the first likelihood calculating means and the second likelihood calculating means. And a means for recognizing the input voice based on the score obtained by the score calculating means, and for each of various voices or noises used for the parameter extraction of the background HMM, Parameter estimating means for obtaining a unique HMM parameter, and the above-mentioned parameter estimating means obtained by the parameter estimating means. Storage means for storing the parameters of each voice or noise HMM, and parameter calculation for obtaining the parameters of the background HMM based on the parameters of each voice or noise HMM stored in this storage means And a second likelihood calculating means, wherein the second likelihood calculating means has the background HM calculated by the parameter calculating means.
A speech recognition apparatus, characterized in that the log likelihood is obtained using M parameters. 5. When changing various voices or noises used for the parameter extraction of the background HMM,
The parameter estimating unit obtains an HMM parameter specific to the voice or noise of the various voices or noises that have not been stored in the storing unit and stores the HMM parameter in the storing unit. Then, the parameter calculating means re-extracts the parameters of the background HMM based on the parameters of each HMM of various voices or noises after the change among the parameters of each HMM stored in the storage means. The voice recognition device according to claim 4, wherein