JPH067352B2 - 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 a Bayesian decision based distance, Mahalanobis distance, and linear discriminant function is applied to a feature parameter of a voice, and an input voice is attempted to be recognized by this. There is.

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, 12 is a coefficient memory, 3 is a discrimination coefficient unit, 6 is a word standard pattern memory, and 7 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 mse.
It is analyzed at every minute time of about c (hereinafter referred to as a frame) and converted into a feature parameter vector. Most of the analysis methods are linear prediction (LPC) analysis methods, and LPC cepstrum coefficients are often used as the characteristic parameters. These parameter vectors are input to the discrimination calculation unit 3. On the other hand, in the coefficient memory 12, various kinds of coefficients necessary for calculating the statistical distance measure by the method described later are stored in advance, and these coefficients are also input to the discrimination calculation unit 3. The discrimination calculator 3 receives the above two inputs, calculates a statistical distance measure, and identifies the phoneme of the corresponding frame. The above process is performed for each frame, and the obtained phoneme sequence is sent to the recognition unit 7. Here, the inter-word distance is comprehensively evaluated between the standard pattern obtained from the word standard pattern memory 6 and the phoneme sequence, 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 And the parameter column vector of the unknown input frame Then, the Bayesian determination corresponds to using the phoneme that maximizes the expression (1) as the identification result. Where n is the order of the vector, The subscript T represents transposition.

Further, the Mahalanobis distance is given by the equation (2), and the phoneme that minimizes the equation (2) becomes 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 discriminant coefficient sequence vector for discriminating between the phoneme u and the phoneme v, and b _u / v is a constant term of the linear discriminant coefficient.

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 scale of the formula (1) or the formula (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.

On the other hand, in the case of the scale of the equation (3), the number of multiplications is the (6)
Although a small amount is required as in the equation, the discrimination result only indicates to which phoneme the unknown input frame relatively belongs, and the distance to the phoneme standard pattern, in other words, the index for the reliability of phoneme identification is obtained. I can't. Therefore, when the recognizing unit 7 in FIG. 2 comprehensively evaluates the phoneme similarity with the word standard pattern 6, the frame with high reliability of phoneme identification 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 reduction of the amount of calculation and possession of a phoneme identification function that can give reliability to an identification result. It is provided.

Means for Solving Problems In order to solve the above problems, the voice recognition device of the present invention is
An acoustic analysis unit that analyzes the input speech for each frame to obtain a characteristic parameter, a first coefficient memory that stores a set of linear discriminant coefficients, a discrimination calculation unit that discriminates phonemes of arbitrary frames, and a predetermined phoneme. A second coefficient memory for storing the feature parameter mean vector and the inverse of the covariance matrix, a Mahalanobis distance calculation unit, a word standard pattern memory for storing a standard phoneme sequence of a word to be recognized, and a word It is configured to have a cognitive function of evaluating the total distance.

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 a feature parameter vector by the acoustic analysis unit and input to the Mahalanobis distance calculation unit and the discrimination calculation unit. The linear discriminant coefficient in the first coefficient memory is also input to the discriminant calculator, where a linear discriminant function is calculated for each pair of phonemes with the feature parameter vector, and the obtained phoneme sequence is the second phoneme string. Is input to the coefficient memory and the recognition unit. Note that here, the calculation amount required for the linear discriminant function calculation is
Only a small amount is required as shown in equation (6). Further, based on the phoneme sequence information input to the second coefficient memory, the corresponding one is selected from the set of the phoneme-specific feature parameter average vector and the inverse matrix of the covariance matrix stored therein. It is input to the Mahalanobis distance calculation unit. The Mahalanobis distance calculation unit receives the above two inputs, the input voice,
A first candidate phoneme distance is calculated from the standard pattern of phonemes to which this belongs. At this time, the distance calculation work only needs to be performed for the only candidate phoneme identified by the discrimination calculation unit, and the overall calculation amount can be expected to be significantly reduced as compared with the conventional example shown in Expression (5). Finally, the recognition unit receives the phoneme sequence from the discrimination calculation unit, the first candidate phoneme distance for each frame from the Mahalanobis distance calculation unit, and the word standard pattern from the word standard pattern memory,
The total distance is evaluated for the entire word 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. The first shows an embodiment of the present invention.

In FIG. 1, 1 is an acoustic analysis unit that analyzes input speech for each frame, 2 is a first coefficient memory that stores a set of linear discriminant coefficients, 3 is a discrimination calculation unit that discriminates phoneme of an arbitrary frame, and 4 is A second coefficient memory for storing a characteristic parameter average vector and an inverse matrix of a covariance matrix that are predetermined for each phoneme, 5 is a Mahalanobis distance calculation unit, and 6 is a standard phoneme sequence of a word to be recognized. A word standard pattern memory, 7 is a recognition unit that evaluates the distance in 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. The analysis method may be any of the conventional methods, but the linear prediction analysis method is used in this embodiment.
If the target voice is limited to the telephone band, it is assumed that both the minimum amount of calculation and the maximum recognition performance are satisfied.
A 10th-order linear prediction analysis is performed at every frame interval of 0 to 20 msec. LPC cepstrum coefficient C _i (i = 1, 2,
..., 10) is desirable. Regarding the LPC cepstrum coefficient, refer to "Digital Processing of Speech Signal" (LR Rabiner RW Chafer, co-authored by "Digi
tal Processing of Speech Signals "). In summary, when the linear prediction model H _(z) is given by the equation (7), C _i is given by the equation (8).

The feature parameter vector {C _i } = (C ₁ , C ₂ , ..., C ₁₀ ) (9) obtained as described above is stored in the discrimination calculation unit 3 in FIG. 1 for each frame. Input and
It is the second input to the Mahalanobis distance calculation unit 5.

In the figure, the discrimination calculation unit 3 performs phoneme discrimination of the input speech for each frame using a linear discriminant function, and inputs the first candidate phoneme obtained as a result to the second coefficient memory 4 and the recognition unit 7. To be done. Since the linear discriminant function is treated in the same manner as described in the section of the prior art in this specification,
The description is omitted here. The first coefficient memory 2 in the figure
Is a linear discriminant coefficient for discriminating between arbitrary phoneme pairs. Are stored, and these coefficients are read out as needed in the discrimination calculation unit 3.

On the other hand, the second coefficient memory 4 is based on the phoneme sequence input from the discrimination calculator 3 And the inverse of the covariance matrix And are selected and used as the first input to the Mahalanobis distance calculation unit 5. The Mahalanobis distance calculation unit 5 receives the first input and the second input, calculates the Mahalanobis distance d _j according to the equation (2), and inputs this to the recognition unit 7 as the first candidate phoneme distance. Where the column vector of equation (2) Each element of C _i is a substitution of each element of C _i in Eq. (9).

By the way, since the distance calculation by the equation (2) is performed for only one phoneme per frame determined by the discrimination calculation unit 3, the number of multiplications required for this is different from the equation (5). That is, since the formula (4) only needs to be calculated once per frame, the number of multiplications required for this is expressed by the following formula (10).

Therefore, the total required multiplication amount in the phoneme identification of the present embodiment is expressed by the following equation (11) per frame. Number (11)
N (J-1) in the equation is the required amount for calculating the linear discriminant function performed by the discriminant calculation section 3 in FIG. 1, and n (n + in the equation (11)
1) is a required amount in the Mahalanobis distance calculation unit 5.

The number of multiplications seen above is summarized in the following table.

The effect of reducing the calculation amount according to the present embodiment is clear from the comparison of the number of multiplications.

Finally, the recognition unit 7 includes the phoneme sequence obtained from the discrimination calculation unit 3, the first candidate phoneme distance obtained from the Mahalanobis distance calculation unit 5, and the standard phoneme sequence of the recognition target word obtained from the word standard pattern memory 6. Using and, the total distance is evaluated for the entire word and the recognition result is output. Although various methods can be considered for the total distance evaluation, the following method is taken as an example of the present invention. That is, when the inter-word distance is accumulated by performing the phoneme level DP matching in frame units using the phoneme sequence of the word standard pattern and the phoneme sequence of the input voice, the representative frame of the input voice on the matching path. When the two phonemes do not match at the position, the smaller the first candidate phoneme distance of the corresponding frame is,
Accumulate heavily weighted distances. 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, the phoneme discrimination for each frame based on the LPC cepstrum coefficient is performed by using the linear discriminant function with a small required calculation amount, and the accuracy of the identified phoneme is unknown only by this. The phoneme distance is calculated by performing the Mahalanobis distance calculation that requires a large amount of calculation only for the phonemes determined to be one candidate phoneme, and thus the reliability of the phoneme identification is also improved during the DP matching at the phoneme level. It is possible to perform distance evaluation in consideration, and as a result, it is possible to realize a voice recognition device having a higher recognition rate than ever before.

EFFECTS OF THE INVENTION As described above, the speech recognition apparatus according to the present invention stores the first set of acoustic analysis means for analyzing the input speech frame by frame to obtain characteristic parameters and the linear discriminant coefficient for phoneme identification. A coefficient memory, a discrimination calculation unit for performing phonological discrimination, a second coefficient memory for storing a predetermined characteristic parameter average vector and an inverse matrix of a covariance matrix for each phonological unit, a Mahalanobis distance calculation unit, and a recognition unit By including a word standard pattern memory that stores a standard phoneme sequence of words to be used and a recognition unit that evaluates the distance over the entire word, the phoneme identification and the reliability evaluation of the phoneme identification can be performed with a small amount of calculation. Both can be done, and as a result, a higher recognition rate than before can be obtained.

[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 ... First coefficient memory, 3 ... Discrimination calculation unit, 4 ... Second coefficient memory, 5 ... Mahalanobis distance calculation unit, 6 ... Word standard pattern memory, 7 ... …
Recognition part.

Claims (1)

[Claims] 1. An acoustic analysis section for analyzing input speech for each frame to obtain a characteristic parameter, and a first set of linear discrimination coefficients stored for the purpose of performing phonological discrimination between arbitrary phoneme pairs. Of the arbitrary frame and the linear discriminant coefficient obtained from the first coefficient memory and the characteristic parameter of the arbitrary frame obtained from the acoustic analysis unit, and the phonological sequence of the discriminated phoneme is output. And a second coefficient memory for storing a feature vector average vector and an inverse matrix of a covariance matrix that are predetermined for each phoneme for the purpose of calculating the Mahalanobis distance between an arbitrary frame and the standard phoneme. , A Mahalanobis distance calculation unit, a word standard pattern memory that stores a standard phoneme sequence of a word to be recognized, and a recognition unit that evaluates the distance in the entire word. Corresponding to the phoneme sequence for each frame obtained from the discrimination calculation unit, a feature parameter average vector for each corresponding phoneme and an inverse matrix of the covariance matrix are selected from the second coefficient memory and are selected for the Mahalanobis distance calculation unit. As a first input, on the other hand, a feature parameter for each frame obtained from the acoustic analysis unit is used as a second input to the Mahalanobis distance calculation unit, and the first input and the second input generated in frames at the same time are used. And a Mahalanobis distance obtained as a first candidate phoneme distance from the Mahalanobis distance calculation unit, and in the recognition unit a phoneme sequence obtained from the discrimination calculation unit and a first Mahalanobis distance calculation unit. Candidate phoneme distance,
A speech recognition apparatus, which performs word recognition using a standard phoneme sequence of words obtained from the word standard pattern memory.