JPH0997096A - Sound model producing method for speech recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To speed up a learning process by forming an initial model and a model to be merged or divided as a single Gaussian distribution model and temporarily producing a mixed Gaussian distribution in the dividing process. SOLUTION: An initial model which has a small signal source is prepared as the initial model and a separating process and a merging process are repeated for the signal source. Further, total likelihood calculated in learning is substituted in P(1) representing the total likelihood when only one signal source is given to generate a model which shares no signal source when the number M of signal sources is 4. Then the size of the output probability distribution of a signal source produced by the merging process is utilized as an evaluation scale to decide the similarity between the signal sources. Then two single Gaussian distributions having two signal sources to be merged are merged into one single Gaussian distribution on the basis of the single Gaussian distribution model and the single Gaussian distribution having a signal source to be divided is re-formed into a two-mixed Gaussian distribution, which is then divided into two single Gaussian distributions.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for generating an acoustic model for speech recognition, and more particularly to a hidden Markov model (Hidden model).
In speech recognition using Markov Model (HMM), the unit of each model, the structure of the state network, the shared structure between multiple states of the signal source for modeling the maximum speech phenomenon with the minimum required model parameters And a method for generating an acoustic model for speech recognition for optimally determining parameters of a signal source.

[0002]

2. Description of the Related Art In order to perform highly accurate and robust speech recognition using an HMM, it is an important issue how to balance model details and robustness. In order to refine the model, it is necessary to properly determine phoneme context categories that cover the entire speech space, and in order to estimate a robust model from a limited number of training speech samples, the model parameters It is necessary to introduce a mechanism that reduces the redundancy of the and efficiently express only the essential information of the voice with the minimum required model parameters.

Due to such a need, the "sequential state division method (Su
ccessive State Splitting (SSS) "has been developed, but the structure of the state network that can be achieved is limited in the processing of only sequential two-division for states, and redundancy of model parameters cannot be completely removed. .

Therefore, the inventor of the present invention filed Japanese Patent Application No. 6-2841.
No. 35, in order to overcome the disadvantage of SSS in which a model is generated only by sequential division into two states, division processing and fusion processing for a signal source are realized at the same time, and one of them is sequentially selected and processed. SSS
It is possible to realize the structure of an arbitrary state network without losing the advantages of the above, and it is possible to express the maximum speech phenomenon with the minimum necessary model parameters with high accuracy and robustness. The method of model generation was provided.

[0005]

[Problems to be Solved by the Invention] However, Japanese Patent Application No. 6-2
The method described in the embodiment of No. 84135 is a method based on a mixture Gaussian distribution model having a mixture number of 2 and sets two two-mixture Gaussian distributions of one set (two) of fusion target signal sources to one. Signal source fusion processing that fuses two two-mixture Gaussian distributions, or one single two-mixture Gaussian distribution that the split-source signal source has
After dividing into two single Gaussian distributions,
The signal source division processing was performed to recreate the Gaussian mixture distribution.

However, it is generally known that learning of a mixed Gaussian distribution model requires much more time than learning of a single Gaussian distribution model, and Japanese Patent Application No. 6-284135. A lot of time had to be spent in the method described in.

Therefore, the present invention realizes the division processing and the fusion processing for the signal source at the same time, and advances the processing while sequentially selecting one of them, so that the advantage of SSS is not lost and an arbitrary state is achieved. It also provides a method for generating an acoustic model for speech recognition with high expression efficiency that can realize the network structure of, and can express the maximum speech phenomenon with high accuracy and robustness with the minimum necessary model parameters. The purpose is to enable fast learning based on a single Gaussian distribution model.

[0008]

[Means for Solving the Problems] As means for achieving the above object, when an initial model and a model to be subjected to fusion processing or division processing are formed as a single Gaussian distribution model and division processing is performed, By temporarily creating a Gaussian mixture with two mixtures, the learning process can be speeded up.

Further, according to the present invention, the fusion and division of the signal sources are performed under the criterion of maximizing the evaluation value with respect to all the learning samples, so that the number of signal sources is locally increased and decreased, and Gradually increases.

As a result, the models are sequentially refined, and finally, the unit of each model, the structure of the state network, the shared structure between a plurality of states of the signal source, and the parameters of the output probability distribution are The acoustic model optimally determined under the common evaluation criteria can be automatically generated at a higher speed than the conventional method.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a flow chart for explaining the outline of the method for generating a voice recognition acoustic model of the present invention. The present invention is directed to a probabilistic model in which the shape of a voice feature pattern (static feature of voice) and its temporal change (dynamic feature of voice) in a minute unit time are expressed as a chain of a plurality of signal sources. Then, by repeating the process of fusing or dividing the individual output probability distributions based on a common evaluation criterion (likelihood maximization), the unit of the model and the structure of the state network, the sharing of the signal source among multiple states The parameters of structure and output probability distribution can be determined simultaneously and automatically.

Hereinafter, a more specific description will be given with reference to FIG. First, a small-scale model (total number of signal sources M = 1 used in the entire model M = 1) is prepared as an initial model (step 1). This is, for example, a model consisting of one state (concept on the model structure associated with a unique phoneme context category) and one signal source (output probability distribution expressed by a single Gaussian distribution and state transition probability). Minimum component) of Then, in the subsequent processing, division and fusion are repeatedly performed on this signal source. Further, by substituting the total likelihood calculated at the time of learning into P (1) representing the total likelihood when the number of signal sources is 1, the model (the number of signal sources M = In 4,
The state forms a model that does not share a signal source).

The model formed during the execution of the method is called Hidden Markov Network (HMnet) and can be represented as a network of multiple states. HMnet is composed of the following information.

(1) Components of HMnet: A set of signal sources. -A set of states. (2) Signal source components: -Signal source number (index). Output probability distribution (single Gaussian distribution in diagonal covariance matrix representation). -Self loop probability and transition probability to the next state. (3) State components: State number (index). -Pointer to the signal source (source number). -Acceptable phoneme environment category (defined as a direct product space of phoneme environment factors). A list of predecessor and successor states.

In the selection of signal sources to be fused (step 3), the size of the output probability distribution of the signal sources generated by the fusion process is used as an evaluation measure in order to judge the similarity between the signal sources. That is, two signal sources Q (i) and Q (j)
The distribution size Dij in the case where the output probability distributions (both are single distributions) are fused for all the combinations of (1) is approximately obtained by Expression (1).

[0016]

[Equation 1]

Two signal sources Q (i) that minimize the value of Dij
And Q (j) are selected as targets of fusion processing. The fusion of the signal sources (step 4) consists of two signal sources Q (i) and Q (i).
This is done by fusing (j) and creating a new signal source Q (I). Mean value μIk and variance σIk of output probability distribution of Q (I)
² can be calculated by the following equations (4) and (5), respectively.

[0018]

[Equation 2]

For the self-transition probability aI ^self of Q (I) and the transition probability aI ^next to the subsequent state, the values obtained by the equations (6) and (7) are used, respectively.

[0020]

(Equation 3)

Q (I) obtained by this processing is Q (I) before fusion.
(i ^) or Q (j ^) is shared in all assigned states. As a process for that purpose, the value is replaced with I for all states in which the value of the pointer to the signal source is i ^ or j ^. By this processing, the number of signal sources in the entire model temporarily becomes M-1.

At this point, it is determined whether or not to use the model obtained as a result of the fusion process for the signal sources. The fusion processing result shows that the total likelihood (which is represented as P ′ (M−1)) obtained from the model after the fusion processing has already been calculated in the processing process before this, and the total distribution number is M−1. Likelihood P at
It will be adopted only when (M-1) is exceeded. In this case, M
The value of is changed to M-1 and the process of model re-learning is proceeded to (step 9). If the result of the fusion processing is not adopted, the phase of the division processing for the model before the fusion processing (the model of step 2) is entered again (to step 5).

Then, prior to the actual division, the signal source to be divided is selected (step 5). For all the signal sources Q (i), the magnitude di of the signal source is calculated by the equation (8), and the signal source having the largest value of di (this is referred to as Q (i ^)) is to be divided. To be selected as.

[0024]

[Equation 4]

Next, Q (i ^) is divided into two signal sources Q (I) and Q (J). As a process for this, first, the Viterbi algorithm is applied to all learning samples using Q (i ^) at the time of likelihood calculation, and the state path of each sample is obtained.

Next, based on the obtained state path, Q (i
Extract all the frames of the learning sample associated with (). After that, the data of all the frames of the extracted learning samples are divided into two groups by vector quantization, and the average value and the variance are obtained for each group. Finally, for each of the two divided signal sources, one of the obtained distributions of each group is assigned as an output probability distribution, and the self-transition probability of Q (i ^) and the value of the transition probability to the subsequent state are set. Copy as it is. Also, the value of M is changed to M + 1.

This processing completes the division of the signal source. When the signal sources are divided, it is necessary to reconstruct the states at the same time. State reconstruction is performed by maximum likelihood PD achieved by recombining only the shared structure of signal sources, maximum likelihood PC achieved when one state is divided in the phoneme environment direction, and one state in the time direction. Among the maximum likelihoods PT achieved in the case of division, the one showing a larger value is adopted.

The rearrangement only of the shared structure of the signal sources (step 6) needs to be performed only when the signal source Q (i ^) to be divided is shared in a plurality of states. . In this case, all subsequent state division processing is continuously performed on the model obtained as a result of the processing here. If Q (i ^) is used in only one state, the processing here is omitted and PD
The value of is set to negative infinity (-∞) and the process proceeds to the next step.

A set of states having a pointer to the signal source Q (i ^) is represented by S. Here, for elements of S, Q
The signal source sharing structure is rearranged by assigning either (I) or Q (J). This allocation is performed by obtaining the maximum value PD calculated by the equation (9).

[0030]

(Equation 5)

At the time when the value of PD is obtained, psI (Ys
)> PsJ (Ys), assign Q (I) to state s, and assign Q (J) to state s otherwise. The state division toward the phoneme environment is performed by dividing one state s among the elements of S into two states and connecting them in parallel.

In this case, it is necessary to distribute the learning sample represented by the route passing through the state to be divided into two routes passing through the newly generated state. This distribution is based on the state s and the phoneme environment factor (factor having two or more elements) f that can be divided in the state s,
This is performed by obtaining a state s ^ and a factor f ^ that maximize PC calculated by equation (10), and dividing the elements belonging to f ^.

[0033]

(Equation 6)

When the state s ^ to be divided and the factor f ^ are obtained, the route to which the element a s ^ f ^ e of f ^ is allocated is determined in the process of calculating the equation (10). The already obtained qI (ys ^ f ^ e) and qJ (ys ^ f ^ e)
) Is used to determine according to equation (11).

[0035]

(Equation 7)

After defining AIf ^ and AJf ^, the state s ^
Is divided into two newly generated states S (I ′) and S (J ′), the following processing is performed. First, I and J are respectively substituted into the pointers to the signal sources in these states. Next, as those phoneme environment information, the factor f ^
AIf ^ and AJf ^ are respectively assigned to the parts related to, and the contents of the factor f held in the state s before the division are copied as they are to the factors f other than f ^. This completes the state division in the phoneme environment direction. The state division in the time direction (step 8) is performed by dividing one state s of the elements of S into two states and connecting them in series. In this case, there are two possibilities depending on which of Q (I) and Q (J) is assigned to the forward state.
Therefore, the application order of the state s ^ and the signal source that maximizes PT calculated by the equation (12) is determined.

[0037]

(Equation 8)

After that, the following processing is performed on the two newly created states S (I ') and S (J') by dividing the state s ^. First, I and J are respectively substituted into the pointers to the signal sources in these states. Next, rI (Y s
^)> RJ (Ys ^), state S (I ') is positioned in the front, otherwise state S (J') is positioned in front,
Reconfigure the network structure. Lastly, as the phoneme environment information, the contents held in the state s ^ before division are copied as they are. This completes the state division in the time direction.

It is highly possible that the HMNet signal source formed at this point is not optimal as a whole model as a result of fusion processing or division processing for some signal sources. Therefore, in order to optimize the parameters of the entire signal source and prepare for the next iterative process, the output probability distribution and state transition probability are calculated for all signal sources within the range affected by the fusion process or the division process. Re-learning (step 9).

After that, the total likelihood achieved as a result of learning is substituted for P (M), and the fusion process and the division process for the signal sources are continued until the number M of signal sources in the entire model reaches a predetermined value. By the processing up to this point, the structure of HMnet is determined. A single Gaussian distribution is assigned to the output probability distribution of each signal source at this point. Therefore, finally, learning (step 10) for changing those output probability distributions into a shape to be finally used is performed on the entire HMnet (when the single Gaussian distribution is used,
This process is unnecessary). This completes the generation of HMnet.

[0041]

According to the acoustic model generation method for speech recognition of the present invention, the fusion and division of the signal sources are repeated while being sequentially selected, so that various speech phenomena can be well represented by the minimum necessary number of signal sources. The effect is that an acoustic model that can be generated can be automatically generated at high speed.

[Brief description of drawings]

FIG. 1 is a flow chart diagram for explaining the mechanism of an embodiment of a method for generating an acoustic model for speech recognition according to the present invention.

[Explanation of symbols]

 1 Initial model creation step 2 Creation step of model example created in the process 3 Selection step of signal sources to be merged 4 Fusion step of signal sources 5 Selection step of signal sources to be divided 6 Recombination step of signal source sharing structure 7 State division step in phoneme context direction 8 State division step in time direction 9 Model re-learning step 10 Distribution shape changing step

Claims (2)

[Claims] 1. A static feature of a voice, which is a shape of a feature pattern of a voice in a minute time, and a dynamic feature of the voice, which is a temporal change thereof, are determined from one output probability distribution and a set of state transition probabilities. A phoneme context-dependent acoustic model generation method for speech information processing using a hidden Markov model modeled as a chain of different signal sources, which is a division or fusion process of signal sources with respect to an initial model having few signal sources. By sequentially selecting, the phoneme context category that is the unit of the model, the number of states used to represent each model and the sharing relationship between multiple models, the sharing relationship of each signal source between multiple states, And the shape of each output probability distribution are all determined under a common evaluation criterion. As a basis, by fusing two single Gaussian distributions of two fusion target signal sources into one single Gaussian distribution, the fusion processing of the signal sources is performed, One Gaussian distribution 2
A method for generating an acoustic model for speech recognition, characterized in that the signal source is divided into two single Gaussian distributions after reforming into a mixed Gaussian distribution. 2. The acoustic model generation method for speech recognition according to claim 1, wherein the division processing for reshaping one single Gaussian distribution of the signal source to be divided into one two-mix Gaussian distribution is a likelihood. The Viterbi algorithm calculates the state path for all learning samples that use the source to be divided during the calculation, and based on the calculated state path, all frames of the learning samples assigned to the source to be divided are extracted. , A method for generating an acoustic model for speech recognition, characterized in that two Gaussian distributions are formed by vector quantization on the data of all frames of the extracted learning samples.