JPH067353B2 - Voice recognizer - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus that performs acoustic analysis and phoneme identification for each frame.

2. Description of the Related Art In recent years, many attempts have been made to improve the performance of a voice recognition device, and a voice recognition device using phoneme identification by a linear discriminant function is one of the targets.

A conventional technique is, for example, Japanese Patent Laid-Open No. 59-13199.
As disclosed in Japanese Patent Publication No. 9, a statistical distance measure such as Bayesian decision based distance, Mahalanobis distance, and linear discriminant function is applied to the feature parameter of the voice, and the input voice is attempted to be recognized by this. There is something.

The conventional speech recognition apparatus as described above will be described below with reference to the drawings.

FIG. 2 shows a conventional voice recognition device. In FIG. 2, 1 is an acoustic analysis unit, 13 is a coefficient memory, 4 is a discrimination calculation unit, 8 is a word standard pattern memory, and 9 is a recognition unit.

The operation of the speech recognition apparatus configured as above will be described below.

The input voice is sent to the acoustic analysis unit 1, where it is 5 to 30 msec.
It is analyzed for each minute time (hereinafter referred to as a frame) and converted into a characteristic parameter. A linear prediction (LPC) analysis method is often used as an analysis method, and an LPC cepstrum coefficient is often used as a feature parameter. These parameters are input to the discrimination calculation unit 3.

On the other hand, in the coefficient memory 13, various coefficients necessary for calculating the statistical distance scale by the method described later are stored in advance, and these coefficients are also input to the discrimination calculation unit 4. The discrimination calculation unit 4 receives the above two inputs, calculates a statistical distance measure, and identifies the phoneme of the corresponding frame. The above-described processing is performed for each frame, and the obtained phoneme sequence is sent to the recognition unit 9. Here, a comprehensive evaluation of the similarity between the standard pattern obtained from the word standard pattern memory 8 and the phoneme sequence is performed, and the word standard pattern closest to the input voice is used as the recognition result of the input voice.

The above-mentioned statistical distance measure is described in the above-mentioned publication as follows.

As a standard pattern for phoneme j, its average The parameter column vector of the input frame Then, the Bayesian determination corresponds to using the phoneme that maximizes the expression (1) as the identification result. Where n is a vector T represents transposition.

Further, the Mahalanobis distance is given by the equation (2), and the phoneme that minimizes the (2) is the discrimination result.

The linear discriminant function is given by the equation (3). If the value on the left side of the equation (3) is positive, the unknown input belongs to the phoneme u, and if it is negative, it belongs to the phoneme v. However, Is a linear discrimination coefficient sequence vector for discriminating between the phoneme u and the phoneme v, and b _u / v is a constant for similarly discriminating between the phoneme u and the phoneme v.

Problems to be Solved by the Invention However, in the above-mentioned configuration, the amount of calculation required to calculate the statistical distance measure is large, or the reliability of the result of phoneme identification is unknown, so that the similarity in the following word units is similar. There were problems such as lack of certainty in the degree evaluation. That is,
When depending on the formula (1) or the scale of (2), The product of the matrix and the matrix must be calculated for all the phonemes j for each frame, and the amount of calculation required for this is enormous, as in equation (5), even if only the number of multiplications is taken.
Therefore, there has been a problem that a device for realizing this is required to be high-speed and large-scale.

When the scale of the above equation (3) is used, the number of multiplications is
Only a small amount is required as shown in equation (6).

However, only the result of the discrimination only indicates which phoneme the unknown input frame belongs to, and cannot obtain the distance from the phoneme standard pattern, in other words, the index regarding the reliability of phoneme identification. Therefore, in the recognizing unit shown in FIG. 2, when the phoneme similarity comprehensive evaluation with the word standard pattern memory 8 is performed, the frame with high reliability and the frame with low reliability of phoneme identification are evaluated with the same weight, and eventually the final frame. There was a problem that the word recognition rate was lowered.

In view of the above problems, the present invention provides a speech recognition apparatus having a high recognition rate, which satisfies the two contradictory requirements of reducing the amount of calculation and possessing a phoneme identification function capable of giving reliability to the identification result. It is a thing.

Means for Solving Problems In order to solve the above problems, the voice recognition device of the present invention is
An acoustic analysis for analyzing the input speech frame by frame, a parameter memory for storing characteristic parameters, a first coefficient memory for storing a set of linear discriminant coefficients, a discriminant calculator for discriminating phoneme of an arbitrary frame, and the phoneme. A phoneme memory that stores the discrimination result, a second coefficient memory that stores a distance coefficient for calculating the phoneme distance, a phoneme distance calculation unit, and a word standard pattern that stores a standard phoneme sequence of words to be recognized. The memory is provided with a recognition unit that evaluates the degree of similarity in the entire word.

Action The present invention has the above-described configuration to solve the above-mentioned conventional problems based on the following actions.

The input speech is converted into characteristic parameters by the acoustic analysis unit and stored in the parameter memory for each frame. Further, the characteristic parameter is input to the discrimination calculation unit. On the other hand, the linear discriminant coefficient in the first coefficient memory is also input to the discriminant calculator,
Here, a linear discriminant function is calculated for each pair of phonemes with the characteristic parameter and stored in the phoneme sequence memory for each frame. Here, a small amount of calculation is required to calculate the linear discriminant function, as shown in the equation (6).

Further, when the same phoneme continuously appears in the phoneme string written in the phoneme string memory, one representative frame is arbitrarily selected from the phonemes and the characteristic parameter having the same frame number as the representative frame is stored in the parameter memory. Read out and input to the phoneme distance calculation unit.

On the other hand, the phoneme information of the representative frame is input to the second coefficient memory, and the applicable one of the distance coefficients for each phoneme stored therein is selected and input to the phoneme distance calculation unit. The phoneme distance calculator receives the above two inputs and calculates the phoneme distance between them. In normal speech, the same phoneme continues for a long time in the vowel stationary part, so that there is only one representative frame for this section, and the calculation by the phoneme distance calculation part does not need to be executed for each frame. For this reason, the total calculation amount can be expected to be significantly reduced as compared with the conventional example.

Finally, the recognition unit receives the phoneme sequence of each frame from the phoneme sequence memory, the phoneme distance of each representative frame from the phoneme distance calculation unit, and the word standard pattern from the word standard pattern memory, The overall similarity evaluation is performed and the recognition result is output.

Embodiment A voice recognition device according to an embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows an embodiment of the present invention. In FIG. 1, 1 is an acoustic analysis unit that analyzes the input speech for each frame, 2 is a parameter memory that stores the characteristic parameters obtained by the acoustic analysis unit 1, and 3 is a first storage unit that stores a set of linear discriminant coefficients. Coefficient memory, 4 is a discrimination calculation unit for discriminating phonemes of arbitrary frames, 5 is a phoneme sequence memory for storing the result of phoneme discrimination, 6 is a second coefficient memory for storing a set of inverse matrix of covariance matrix for each phoneme, Reference numeral 7 is a Mahalanobis distance calculation unit, 8 is a word standard pattern memory that stores a standard phoneme sequence of words to be recognized, and 9 is a recognition unit that evaluates the degree of similarity of the entire word.

The operation of the speech recognition apparatus configured as described above will be described below.

First, the input voice is analyzed by the acoustic analysis unit 1. Although any conventional analysis method may be used, a linear prediction analysis method is used in this embodiment. Also, if the target voice is limited to the telephone band, it is assumed that both of the minimum calculation amount and the maximum recognition performance are satisfied, 8 KHz, 1
After performing 2-bit sample quantization, 10th-order linear prediction analysis is performed at every frame interval of 10 to 20 msec. LP
It is desirable to obtain the C cepstrum coefficient (C _i , i = 1, 2 ... 10). Regarding the LPC cepstrum coefficient, refer to "Digital Processing of Speech Signal" (L.
R.Rabiner, RWSchafer, “Digital Processing of
Speech Signals ”). In summary, when the linear prediction model H _(z) is given by the equation (1), the LPC cepstrum coefficient C _i is given by the equation (8).

The characteristic parameters obtained as described above are stored for each frame in the parameter memory 2 in FIG.

The discriminant calculation unit 4 uses the linear discriminant function to identify the phoneme of the input voice for each frame, and sequentially writes the results in the phoneme sequence memory 5. The linear discriminant function has been described in the section of the prior art in this specification, and is omitted here. Further, in the figure, the first coefficient memory 3 stores linear discriminant coefficients for discriminating between arbitrary phoneme pairs, and these coefficients are read out as needed in the discriminant calculation section 4. .

By the way, normal voice is 10 to 20 ms shown in this embodiment.
When the phoneme is identified in frames of about ec, it is usual that the same phoneme is continuously identified and output in a plurality of frames due to the steadiness of the acoustic characteristics of the vowel part. In the present embodiment, one representative frame is arbitrarily selected from among them, and the characteristic parameter having the same frame number as that is read from the parameter memory 2 and input to the Mahalanobis distance calculation unit 7.

On the other hand, the phoneme information in the representative frame is stored in the phoneme sequence memory 5
Is read out from the above and input to the second coefficient memory 6, and the corresponding one is selected from the phoneme-based inverse matrix stored therein, and becomes the first input to the Mahalanobis distance calculation unit 7. . On the other hand, the frame number information of the representative frame is also input to the parameter memory 2, and the corresponding feature parameter becomes the second input to the Mahalanobis distance calculation unit 7. The Mahalanobis distance calculation unit 7 receives the first and second inputs, calculates the Mahalanobis distance according to the equation (2), and outputs the result to the recognition unit 9. As described above, since only one representative frame can be obtained for consecutive identical phoneme intervals, the calculation by the Mahalanobis distance calculation unit 7 is significantly reduced as compared with that described in the conventional example.

Finally, the recognition unit 9 receives the phoneme sequence of each frame from the phoneme sequence memory 5, the Mahalanobis distance of each representative frame from the Mahalanobis distance calculation unit 7, and the word standard pattern memory 8
It receives the standard pattern from the word and evaluates the overall similarity of the whole word, and outputs the recognition result. Although various methods can be considered for the comprehensive similarity evaluation, the following method is taken as an example of the present invention. That is, when the phoneme level DP matching is performed in frame units using the word standard pattern and the phoneme sequence to accumulate the inter-word distances, the phoneme of the input voice at the representative frame position of the input voice is detected on the matching path. When the phonemes of the standard patterns do not match, the smaller the Mahalanobis distance is, the larger the weighted distance is accumulated. The smaller the Mahalanobis distance is, the more reliable the phoneme identification result in the corresponding frame is.Therefore, it is a reasonable method that the disagreement of the phoneme class there has a greater effect on the increase in the distance for the whole word. is there.

As described above, according to the present embodiment, phoneme discrimination for each frame by the LPC cepstrum coefficient is performed by using a linear discriminant function having a relatively small calculation amount, and a large amount of calculation is performed only for the representative frame selected as the phoneme center. It is possible to calculate the Mahalanobis distance, which requires a large amount, and to calculate the phoneme similarity that cannot be obtained only by the above linear discriminant function. The phoneme distance information in the representative and quantitative representative frame can be globally and quantitatively input as the DP matching operation over the entire length of the word, and as a result, higher than the conventional speech recognition device. A voice recognition device capable of obtaining a recognition rate can be realized.

EFFECTS OF THE INVENTION As described above, according to the present invention, the acoustic analysis unit that analyzes the input speech for each frame, the parameter memory that stores the characteristic parameter, the first coefficient memory that stores the set of linear discriminant coefficients, and the arbitrary frame. A discriminant calculation unit that discriminates the phoneme of
A phoneme memory that stores a phoneme discrimination result, a second coefficient memory that stores a distance coefficient for calculating a phoneme distance, a phoneme distance calculation unit, and a word that stores a standard phoneme sequence of words to be recognized. By providing a standard pattern memory and a recognition unit that evaluates the degree of similarity in the entire word, there are two contradictory requirements of reducing the amount of calculation and possessing a phoneme identification function that can give reliability to the identification result. Filled,
A voice recognition device having a high recognition rate can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention, and FIG. 2 is a block diagram of a conventional voice recognition device. 1 ... Acoustic analysis unit, 2 ... Parameter memory, 3 ... First coefficient memory, 4 ... Discrimination calculation unit, 5 ... Phoneme string memory, 6 ... Second coefficient memory, 7 ... Mahalanobis distance Calculation unit, 8 ... Word standard pattern memory, 7 ... Recognition unit.

Claims (2)

[Claims] 1. An acoustic analysis unit for analyzing input speech frame by frame, a parameter memory for storing characteristic parameters obtained by the acoustic analysis unit, and a predetermined unit for performing phoneme discrimination between arbitrary phoneme pairs. Phoneme discrimination of the arbitrary frame using a first coefficient memory for storing a set of linear discrimination coefficients, a characteristic parameter of the arbitrary frame obtained from the parameter memory, and a linear discrimination coefficient obtained from the first coefficient memory. And a phoneme sequence memory that stores the result of the judgment by the judgment calculator, and a predetermined distance coefficient for each phoneme for calculating the phoneme distance between an arbitrary frame and the standard phoneme. A second coefficient memory, a phoneme distance calculator that calculates a phoneme distance between an arbitrary frame and a standard phoneme, and a standard phoneme sequence of words to be recognized. And word standard pattern memory, comprising a recognition unit for evaluating the similarity of the entire word, when the same phoneme in the phoneme sequence memory is written sequentially,
The characteristic parameter having the same frame number as the arbitrarily selected representative frame is selected from the parameter memory, and the characteristic parameter corresponding to the phoneme discriminated by the discrimination calculation unit is selected from the second coefficient memory. Select a distance coefficient for each phoneme, calculate the representative phoneme distance in the representative frame by the two selected as described above,
Character recognition is performed using a phoneme string obtained from the phoneme string memory, a representative phoneme distance obtained from the phoneme distance calculation unit, and a standard phoneme string of words obtained from the word standard pattern memory. And a voice recognition device. 2. A second coefficient memory stores an inverse matrix set of inverse matrices of phoneme-specific covariance matrices that are predetermined for each phoneme in order to calculate a Mahalanobis distance between an arbitrary frame and a standard phoneme. The speech recognition apparatus according to claim 1, wherein the speech recognition device is a memory, and the phoneme distance calculation unit is a Mahalanobis distance calculation unit.