JP2001083986A - Method for forming statistical model - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the confirmation ratio and shorten the HMM forming time by determining the tolerance when the same category as learning data is imparted to the learning data, selecting the tolerance of this learning data as a judgment reference, and learning a statistical model by use of only the selected learning data. SOLUTION: 'Learning data 1' for learning an acoustic model 1 (HMM1)' used for determining the tolerance and 'learning data 2' to be selected, which is the data for learning an 'acoustic model 2 (HMM2)' for recognition, are prepared. A preliminarily HMM1 ('acoustic model 1') is formed by use of the 'learning data 1'. The preliminary HMM1 formed therein is used to determine the tolerance when a right answer (for example, right answer sound element sequence) is imparted to each data of the 'learning data 2'. A threshold is provided in this tolerance, and a 'leaning data 2'' having a tolerance exceeding the threshold is selected to generate the 'acoustic model 2' for recognizing processing.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical model,
For example, in pattern recognition expressed using a Hidden Markov Model (HMM) (for example, Nakagawa et al., "Speech Recognition in Stochastic Models", IEICE, 1997), likelihood is used at the time of model creation. A robust and high-performance model is created by selecting the learning data at the time of model creation by using the statistical model.
The present invention relates to a method for creating a statistical model that can shorten the creation time.

[0002]

2. Description of the Related Art Although the present invention has been described using voice as an example,
In recognizing which category the input voice data belongs to, the characteristics of the input voice data are represented by a statistical model, and the statistical model (hereinafter, referred to as HMM for more specific description, The present invention is applicable to a model created when performing various pattern recognition using any statistical (probability) model, for example, a neural network.

In speech recognition, an HMM (phoneme model, syllable model, word model, etc.) obtained from learning speech data is used.
Is compared with the input voice data, the degree of matching between the two is determined as likelihood, and a recognition result is obtained. The parameters of the HMM are the conditions for recording the learning speech data (background noise, line distortion,
Greatly depends on the speaker. Therefore, if the voice recording conditions are different from the actual recognition time, a mismatch between the input voice pattern and the HMM occurs, and as a result, the recognition rate decreases.

[0004]

SUMMARY OF THE INVENTION Input voice data and HM
In order to prevent a reduction in the recognition rate due to mismatch with M, a model may be created using audio data recorded under the same conditions as those for performing recognition. However, a model based on a statistical method such as an HMM requires a long time for the creation process (about 400 hours or more). Further, among these audio data, there is included audio data having characteristics that are significantly different from audio data to be evaluated (hereinafter referred to as evaluation data). When such speech is used for learning, the accuracy of the HMM may be degraded.

[0005] The present invention has been made in view of the above,
The purpose is to improve the recognition rate by calculating the likelihood of the learning data when learning the HMM and providing a method of selecting the learning data based on the likelihood, and improving the HMM creation time. Is to shorten.

[0006]

In order to achieve the above object, according to the first aspect of the present invention, a statistical model expressing the features of each recognition category is used for an input vector time series corresponding to input data. In pattern recognition, which calculates the likelihood for each recognition category and outputs the category represented by the statistical model with the highest likelihood as a recognition result, the learning data used when creating the model is calculated by using preliminary statistics to determine the likelihood of the learning data. A model is created, using this preliminary statistical model, a likelihood is obtained when the same category as the training data is given to the training data, and the likelihood of the training data is selected as a criterion. Learn the statistical model using only

In the first aspect of the present invention, the likelihood for the correct answer of the learning data is obtained by using the likelihood as the criterion at the time of recognition, and a threshold is provided to select the learning data.

According to the second aspect of the present invention, the likelihood value is used as a threshold as a criterion for selecting learning data, and input data whose likelihood is equal to or larger than the threshold are selected.

According to the third aspect of the present invention, input data is arranged in order of likelihood based on the selection criteria of learning data.
Use higher-order data.

[0010]

FIG. 1A shows an example of a three-state HMM. Such a model is created for each voice unit (category). To the states S1 to S3, statistical distributions D1 to D3 of the audio feature parameters are given, respectively. For example, assuming that this is a phoneme model, the first state represents the statistical distribution of the feature amount near the beginning of the phoneme, the second state near the center, and the third state near the end.

The feature distribution of each state of the HMM is represented by a plurality of continuous probability distributions (hereinafter, referred to as “multiple probability distributions”) in order to express a complicated distribution shape.
(Referred to as a mixture continuous distribution) in many cases. Although various distributions can be considered as the continuous probability distribution, a normal distribution is often used. Further, each normal distribution is often represented by a multidimensional uncorrelated normal distribution having the same number of dimensions as the feature amount.

FIG. 1B shows an example of a continuous mixture distribution. ₁ mean vector mu In this figure, a normal distribution N (μ _{1, σ 1)} variance is sigma ₁ and similar N (μ _2, σ ₂₎ and the same N (μ _3, σ ₃₎ and Are shown as three normal parts. For example, a distribution such as distribution D1 is shown in FIG.
Are represented as three normal components.

The output probability b () in the state of the mixed continuous distribution HMM with respect to the input feature vector X _t = (X _{t, 1,} X _{t, 2,} ... X _{t, p} ) ^T (P is the total number of dimensions) at time _t X
_t )

[0014]

(Equation 1)

The calculation is as follows. Here, W _k represents a weight coefficient for the k-th multidimensional normal distribution k included in the state. Probability density P _k (X _t ) for multidimensional normal distribution k
Is

[0016]

(Equation 2)

Is calculated as follows. Where μ _k is the mean vector for the k-th multidimensional normal distribution k of the state, Σ
_k also represents a covariance matrix. If the covariance matrix is only diagonal components, that is, a diagonal covariance matrix, then P
The logarithmic value of _k (X _t ) is

[0018]

(Equation 3)

## EQU1 ##

Here, μ _{k, i} is the ith component of the mean vector of the k-th multidimensional normal distribution of the state _, and σ _{k, i} is the covariance of the k-th multidimensional normal distribution of the state. Represents the ith diagonal component (variance value) of the matrix. Thus, the likelihood is HM
M and the similarity of audio data.

FIG. 2 shows a flow chart for creating an acoustic model. The audio data is analyzed by the acoustic analysis unit, and the HMM calculates the above W, μ, σ using the correct answer label corresponding to the audio.
Is estimated from the initial acoustic model.

As audio data, for example, "all realities
When data such as ".." or "video game or..." Is given, "arayuru..." Or "te
lebi ... "is prepared. And, for example, "a"
, "R" and so on, an initial acoustic model is prepared as a so-called sample of an acoustic model to be obtained as a result of learning.

The illustrated initial acoustic model is, for example, corresponding to “a”, is a 26-dimensional, 4-mixed, 3-state model, and has a total of 312 normal distributions N (μ ₁ , σ ₁ ), N (μ ₂ , σ ₂ )... To specify N (μ ₃₁₂ , σ ₃₁₂ ) μ ₁ (mean) 0.0; σ ₁ (variance) 1.0 μ ₂ (mean) 0.0; σ ₂ (Dispersion) 1.0 ... is indicated. The learning unit illustrated includes (i) a linear prediction analysis (LPC) cepstrum or another MFCC (mel cepstrum) obtained by analyzing each sound data by the acoustic analysis unit, (ii) a correct answer label, iii) Given the initial acoustic model and learning results, for example, “a” is μ ₁ (average) 0.01; σ ₁ (variance) 0.2 μ ₂ (average) −0.03; σ ₂ (Variance) 0.04... Is represented by a composite of a plurality of normal distributions.

In the subsequent recognition processing, for example, “a” in the input data to be recognized is obtained in the same acoustic model as described above, and then the correct acoustic model for “a” is obtained. The distance is calculated to recognize that “a” in the input data corresponds to the correct answer “a”.

The learning process in the illustrated learning unit includes a conventionally known (i) learning algorithm (discrete HMM) (ii) learning algorithm (multidimensional normal distribution) (iii) learning algorithm (mixed normal distribution) (iv) learning algorithm ( Semi-continuous HMM) can be used. The algorithms are described by Kiyohiro Kano, Satoshi Nakamura and Shiro Ise, published by Kuniaki Ai, published by Shokodo Co., Ltd. November 1, 1997
This is described on pages 74 to 79 of "Digital Signal Processing of Voice / Sound Information (Digital Signal Processing Series 5)" issued on the first edition of the 0th edition, and the inventor described the above "learning algorithm ( Mixed normal distribution) ".

FIG. 3 shows a flowchart of the recognition. In the recognition, as shown in FIG. 3, for the recognition candidate model, likelihood calculation is performed on the feature vector of each frame of the input speech, and the logarithmic value of the likelihood obtained by dividing the cumulative total of the obtained speech by the number of frames. A recognition result is output using a word dictionary in which the word and a recognition candidate are described.

That is, the “acoustic model” obtained in FIG. 2 is used as the “acoustic model” shown in FIG. 3, and the evaluation data (that is, the input data to be recognized), for example, Are not provided "," Japan-US relations will be important ",..., Etc., and the above-mentioned LPC cepstrum is obtained by the acoustic analysis unit, and then supplied to the recognition unit.

As described above, after the acoustic model is prepared, the recognition process is performed in association with the "word dictionary". According to the present invention, a "acoustic model" more preferable than the prior art is obtained.

In the recognition section, the LP from the acoustic analysis section
The distance between the acoustic model obtained using the C cepstrum and the “acoustic model” as shown in FIG. 2 described above is calculated, and for example, “d”, “e”, “m”, “o”, “y”, “a”,.・ Recognition is obtained and “recognition result”
Get.

As described above, after the acoustic model is prepared, the recognition process is performed in association with the "word dictionary". In the present invention, a more preferable "acoustic model" can be obtained.

That is, in the present invention, the likelihood for the correct answer of the learning data is obtained by using the likelihood which is a criterion at the time of recognition, and a threshold is provided to select the learning data. By using the likelihood as a criterion in this way, it is possible to exclude data having an incorrect correct label, data containing noise, data in which voice is cut off halfway, and the like.

FIG. 4 shows a flowchart of the learning data selection based on the likelihood. "Learning data 1" for learning "Acoustic model 1 (HMM1)" used for obtaining likelihood and "Learning model 2 (HMM2)" for recognition are data to be selected. Learning data 2 "
And prepare. At this time, the learning data 2 may or may not include the learning data 1.

First, the preliminary HMM 1 is set using the learning data 1.
(“Acoustic model 1”) is created. Using the preliminary HMM 1 created here, the likelihood when a correct answer (for example, a correct phoneme sequence) is given to each data of the learning data 2 is obtained. Next, a threshold value is set for this likelihood, and learning data 2 ′ having a likelihood greater than or equal to the threshold value is selected, and an acoustic model 2 for recognition processing is selected.
Generate

FIG. 5 shows a flowchart for selecting learning data based on the likelihood threshold. In FIG. 5, “learning data 2” includes “prevention and health management rehabilitation
.. "and" No exit ... "are given arbitrarily. Then, the LPC cepstrum is obtained by the “acoustic analysis unit”, and “acoustic model 1 (“ acoustic model ”in FIG. 2);
"Acoustic model 1" in Fig. 4) "
The likelihood is calculated by the “likelihood calculation unit” using the (correct answer label as shown in FIG. 2).

Based on the result, the learning data selection unit arranges the learning data input as "learning data 2" in order from the one having the highest likelihood. In the case shown in the figure, No. 3 “Take a small amount of income” got likelihood “79.6”, No. 5 got likelihood “77.7”, and No. 2 “Exit Absent··
.. Has a likelihood of “74.8”, it is determined that “learning data 2 ′” (“learning data 2 ′” in FIG. 4) is No. 3 and No. 5 , No.2,
No. 6, No. 1, and No. 7 are each selected as having a likelihood of “70” or more.

FIG. 6 shows a flowchart for selecting the learning data based on the upper x% of the likelihood. The learning data of the upper few% to several tens% is selected, and learning data 2 'is created. That is, in the case of FIG. 6, the uppermost x% of the learning data is selected from those having the highest likelihood. In the case shown in the figure, No. 3, No. 5, No. 2, No. 6, and No. 1 are selected as the top x%. That is, it is selected as "learning data 2 '".

The "learning data 2 '" is selected by any of the above-mentioned methods, and FIG.
As shown in (1), “acoustic model 2 (HMM2)” is created.

At the time of this learning, the amount of the learning data 2 'can be adjusted by setting the threshold value, and the time required for creating the HMM 2 can be adjusted.

A speech recognition experiment was performed to see the effect of learning data selection. The machine used to create the HMM is a Sun Ultra Enterprise 450 MHz. The learning data uses a news broadcast voice, and the evaluation recognizes 50 sentences of the news voice. Vocabulary size is 20,000
Is a word. Experiment results

[0040]

[Table 1]

## EQU1 ## The learning data 1 is not selected, the recognition rate of the model learned over 300 hours using 6666 sentences is 93.23%, and the learning data selection according to the present invention is performed over 240 hours using 3894 sentences. The recognition rate of the model thus obtained was 93.79%. By performing the learning data selection according to the present invention as described above, the HMM
Creation time was reduced by 60 hours and the recognition rate was 0.56.
% Improvement was seen.

[0042]

As described above, according to the present invention, H
In the case of creating an MM, the recognition rate is improved by selecting the learning data particularly by using the likelihood, and the HM is selected.
The time for creating M can be reduced. In general, a specific speaker and a voice in a uttered state can be used as the learning data. When speech recognition is performed using an acoustic model created by such speech, the recognition rate of speech having characteristics different from those at the time of learning is significantly reduced. However, according to the present invention, the likelihood with a model that has been learned in advance is calculated, and it is possible to avoid using speech having an abnormal value as the learning data. Therefore, it is possible to suppress a decrease in the recognition rate using the created acoustic model.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating an HMM.

FIG. 2 shows a flow for creating an acoustic model.

FIG. 3 shows a flow of a recognition process.

FIG. 4 shows a flow for selecting learning data according to the present invention.

FIG. 5 shows an example of using a likelihood threshold in selecting learning data.

FIG. 6 shows an example in which the upper x% of the likelihood is adopted when selecting learning data.

 ────────────────────────────────────────────────── ─── Continued on the front page (72) Inventor Shoichi Matsunaga 2-3-1 Otemachi, Chiyoda-ku, Tokyo Nippon Telegraph and Telephone Corporation F-term (reference) 5D015 GG04 HH03 HH04 HH14

Claims (3)

[Claims] An input vector time series corresponding to input data is used to calculate a likelihood for each recognition category by using a statistical model expressing features of each recognition category, and a statistical model having the highest likelihood is represented. In pattern recognition, which outputs categories to be recognized as recognition results, a preliminary statistical model for calculating the likelihood of learning data is created for learning data used in model creation. A statistical model creation method characterized in that the likelihood in the case where the same category is given is obtained, the likelihood of the learning data is selected as a criterion, and a statistical model is learned using only the selected learning data. 2. The statistical model creating method according to claim 1, wherein a likelihood value is used as a threshold as a criterion for selecting learning data, and input data having a likelihood equal to or larger than the threshold is used. 3. The statistical model creation method according to claim 1, wherein input data is arranged in order of likelihood as a criterion for selecting learning data, and higher-order data is used.