JPH0634186B2 - Speech recognition method and apparatus thereof - Google Patents

Description

TECHNICAL FIELD The present invention relates to a large vocabulary speech recognition method using vector quantization and an apparatus thereof.

2. Description of the Related Art FIG. 1 is a basic circuit diagram of a voice recognition device. In the figure, 1 is a microphone, 2 is an analysis unit, 3 is a changeover switch, 4 is a standard pattern unit, 5 is an input voice pattern unit, and 6 is distance calculation. Reference numeral 7 denotes a minimum value detecting portion, 8 denotes a recognition result portion, and the distance calculating portion 6 and the minimum value detecting portion 7 form a pattern matching portion. In FIG. 1, first, a voice coming from the microphone 1 is analyzed to extract a pattern for recognizing the features of the voice pattern. In the system for a specific speaker, the analysis result of each recognition target word of the speaker is registered in advance as a standard pattern before recognition, and at the time of recognition, the standard pattern of each recognition image word and the input voice pattern are recognized. And the closest recognition word is selected, that is, the recognition target word having the smallest distance is selected. In the case of an unspecified speaker, a standard pattern that can absorb individual differences is used.

FIG. 2 is a diagram showing an example of an analysis method using a bandpass filter group (BPF), which is “3” (/ san /).
Is a time change diagram of a spectrum pattern obtained by analyzing (BPF analysis) the following voice with a bandpass filter group of 16 channels (all bands are 200 to 6000 Hz). One unit of the time axis is 18 ms, and if you take a cross section at a certain time, it becomes the spectrum at that time. Actual recognition processing is all digital processing, and the spectrum of a horizontal row at a certain time i The intensity value is represented by the feature vector ai (= ai ₁
ai ₂ ai ₃ ... ai ₈ ... ai ₁₆ ), and the input voice pattern (here, the voice pattern of “3”) is A = a ₁
a ₂ ... A _i ... A _I (I = 32).

Therefore, the voice pattern is expressed as follows.

A = a ₁ a ₂ ...... a _i ...... a _I (1) ai is a quantity representing the feature of the voice at time i, and is generally a vector value, and A is the feature vector ai [i = 1.
~ 32 (when I = 32)], and I corresponds to the length of the voice pattern A.

Further, the vector ai is called a feature vector and is represented by ai = (ai ₁ , ai ₂ ... ai _q ... ai _Q ) (2). Q is the order of the vector, which in the example of FIG. 2 corresponds to the channel number 16 of the bandpass bandpass filter group.

Likewise the standard pattern of the word n and ^{B n,} represented by _{^{B n = b 1 n b 2}} n ... b j n ... b Jn n ... (3). At this time, b _j ⁿ is a feature vector at the time j of the standard pattern of the word n, and has the same degree as the feature vector ai of the input pattern A. Further, J ⁿ represents the length of the standard pattern of the word n, n is a serial number indicating the word name, and Σ is a recognition word set of N words, where Σ = {n | n = 1,2 ... n } (4) However, if there is no need to specify a particular word is omitted subscripts _{n, B = b 1 b 2} ... b j ... b J ...... (5) bj = bj 1, bj 2, ... bj 8 ... bj _Q ) ... (6)

In the voice recognition process, the standard pattern B ⁿ of all the words in the recognized word set for the input pattern A is subjected to pattern matching while time-normalizing, and the word n closest to the input pattern A is searched from the N words.

Figure 3 is a mapping model for time normalization, which is
In the above example, the standard pattern B of the word "3" is aligned with the time axis of the input pattern by the mapping function.
Usually, the mapping function is expressed by j = j (i) (7), which is called a distortion function.

If this distortion function is known, the time axis of the standard pattern B is converted by the equation (7) to obtain the time axis i of the input pattern A.
However, in practice, this distortion function is unknown, so that one pattern is artificially distorted so that it is most similar to the other, i.e., the distance is minimized and the optimal distortion is obtained. I try to define the function.

FIG. 4 is a diagram for explaining an example of the DP matching method for implementing the above principle. Now, considering the distortion function j (i) as a function for distorting the time axis of the standard pattern B, this distortion is considered. The pattern B is converted into the following pattern B ′ by the function j (i).

B '= bj ( ₁ ) bj ( ₂ ) ... bj ( _i ) ... b
j ( _I ) ... (8) In consideration of the time distortion development of the actual voice pattern, for example, (a) and j (i) are (approximately) a monotonically increasing function, and (b) , J (i) is (approximately) a continuous function, (c), j (i) takes a value near i, etc., but there are almost infinite distortion functions that satisfy these conditions. However, the distortion function j (i) is determined so that B ′ is the most similar to the input pattern A, that is, the distance is the smallest. For this purpose, first, the time axis of the standard pattern B is set to the distortion pattern j (i) in the input pattern A
The pattern B'is obtained by mapping on the i-axis of the above. At this time, the distortion function j (i) that minimizes the distance between the pattern A and the pattern B'is the optimum distortion function. This input pattern A
And the mapping pattern B ′ is It is represented by. Here, ‖ ‖ indicates the distance between two vectors. Then, the problem of minimizing the distance in the above equation (9) is Is defined by Generally, D (A, B) is called a time-normalized distance or a distance between patterns, and d (i, j) is a vector a.
The distance between i and bj is usually called the inter-vector distance.

In FIG. 5, the (i, j) plane shown in FIG. 4 is abstracted into a grid plane, and the coordinates (i, j) of each grid point are represented.
When the above equation (10) is considered on this plane, the optimum distance from (1,1) to (I, J) is obtained. I will search for a route, but in this case i-1
In many cases, the number of paths from the state 1 to the state i is limited to three as shown. The matching window is for preventing extreme time distortion, and serves as the matching window and satisfies the above three conditions (a) to (c) regarding the time normalization. Now, each i of i = 1, 2 ... I
Considering a case in which the state j to be moved next is optimally controlled to minimize the evaluation function of the expression (10), the initial condition is g (1,1) = d (1 , 1) (12) The recurrence formula is The distance between patterns is D (A, B) = g (I, J) (14), and the calculation of the equation (13) is performed in the direction of increasing (i, j) at the grid points in FIG. Will be followed.
That is, g (i, j) is (i, j) from the (1,1) point.
The minimum sum of distances to reach the point.
The expression is g (i-1, j), g (i-) that has already been obtained for j, (j-1), (j-2) in the (i-1) th stage.
1, j-1), g (i-1, j-2), g (i, j) in the state j at the i-th stage is obtained.

FIG. 6 is a block diagram of a processor that executes the above-mentioned DP (Dynamic Programmig) matching processing. In the figure, 11 is A memory, 12 is B memory, and 13 is d (i, j) calculation. Part, 14 is a g (i, j) calculation part, 15
Is a G (j) memory, 16 is a control unit, the d (i, j) calculation unit 13 calculates the inter-vector distance between ai and bi, and g
(I, j) The shortest distance g to reach (i, j) in the calculation unit 14
Calculate (i, j) and process them in parallel. g (i,
j); g is stored in the D (j) memory 15 when j = I to J is calculated.
(I-1, j); j = 1 to J are included. Also, mi
n detects the smaller of g ₁ and g ₂ and sets the smaller value to g
Put in.

When the DP matching method is used, the (13)
As is clear from the first term on the right side of the equation, if the matching window is not provided, at least I × J × N (N is the number of registered words) calculations are required.

In order to reduce the distance calculation amount by the DP method, a split method taking an onomatopoeic unit has been proposed. In this spirit method, the distance calculation for each frame of the input speech is limited to a predetermined number (K). ) Onomatopoeia (codebook)
Just go to and save it in the form of a matrix, and
In P matching, the amount of calculation of the distance is reduced by simply searching the matrix. Only the word standard pattern is subjected to vector quantization by the split method, and the vector quantization is not applied to the input voice. Thus, in this spirit method, a distance matrix between the analysis frame of the input voice and the pseudophony (vector) stored in advance is created. In this distance matrix, the horizontal axis is the frame number of the input voice and the vertical axis is Is an onomatopoeia (vector) number, and DP matching between the standard pattern stored as a vector number series and the input voice is performed with reference to this distance matrix.

A double split method has been proposed as a further improvement of the above split method. This double split method is a method of vector quantizing not only a standard pattern but also an input voice.

FIG. 7 is a block diagram for explaining an example of the double split method. In the figure, 20 is an input unit, 21 is an analysis unit, 22 is a vector quantization unit, 23 is a vector number generation unit, and 24 is a vector number generation unit. A standard vector storage unit, 25 is a vector distance matrix table, 26 is a DP matching unit, 29 is a minimum distance word identification unit, 30 is a recognition result collection output unit, and input speech is stored in the standard vector storage unit 24 in the vector quantization unit 22. It is converted into a standard feature vector and quantized, converted into a vector number sequence in the vector number generation unit 23, and sent to the DP matching unit 26. In the DP matching section 26, the D of the vector number series of the input voice sent as described above and the vector number series of the word standard pattern stored in the vector storage section 24 in advance.
The P matching is executed with reference to the inter-vector distance matrix table 25 representing the inter-vector distance, and the word having the minimum distance is determined by the word identifying unit 29.
The result is output to the recognition result output unit 30.

More specifically, the feature vector ai of the i-th frame of the input voice A is ai = (ai ₁ , ai ₂ , ai _p ) where i = 1, 2, ... I (I; number of input frames) P; represented by the number of dimensions of the feature parameter, while k of the standard vector pattern B of the voice
The th feature vector bk is represented by bk = (bk ₁ , bk ₂ ... bk _p ) where k = 1, 2, ... K (K; the number of quantized standard vectors) (this feature vector bk is vector quantized). Has been).

Here, in order to perform vector quantization on the feature vector ai of the input voice, the feature vector ai is compared with the standard pattern in the standard vector storage unit. Convert to. However, Is. However, Is the value of k that minimizes | (ai-bk) |. That is, the distance between ai and all bk is calculated, and the feature vector whose distance is closest to ai Instead of ai, It is what Here, when the inter-frame distance between bk and bj (k ≠ j) is calculated in advance and stored in the vector distance matrix table 25, the feature vector ai of each input frame and the vector quantized word standard pattern bk The inter-frame distance between and is obtained by referring to the table 25. Here, the standard pattern of registered words is obtained by a vector number series of standard vectors.

The double split method (1) can reduce the distance calculation amount as compared with the split method.

(2) Since the distance matrix can be set in advance, the distance matrix can be skillfully set.

(3) LPC by using a simple scale for input speech
The calculation of the characteristic parameters in the analysis can be omitted.

And the like, and is particularly useful for specific speaker recognition. However, even in the case of feature vectors corresponding to the same phoneme in the same word, there is a large difference between speakers. Therefore,
A plurality of standard patterns are often used in speech recognition of an unspecified speaker (multi-template method, KNN method, etc.).
However, the amount of calculation and the amount of memory increase in proportion to the number of standard patterns, which is a practical problem.

Purpose The present invention has been made in view of the above-mentioned circumstances,
In particular, the speech recognition method and apparatus using the double split method are aimed at improving recognition accuracy and recognition speed.

Configuration In order to achieve the above-mentioned object, (1) the input voice is quantized, the quantized input voice is indexed, and a vector-to-vector distance of a class corresponding to the standard pattern is determined by a class number sequence of the standard pattern. Selecting a matrix table and referring to the vector number sequence of the standard pattern to perform DP matching with the input voice pattern, or
(2) Quantizing means for quantizing the input speech, indexing means for indexing the input speech quantized by the quantizing means, quantized vector number sequence and class number sequence classified into classes A standard pattern consisting of, a vector-to-vector distance matrix table in which the vector-to-vector distance between the standard pattern and the input speech quantized by the quantizing means is pre-classified according to the type of phoneme, and the standard pattern The present invention is characterized by including DP matching means for selecting the inter-vector distance matrix table corresponding to the attached class number sequence and matching the vector number sequence of the standard pattern with the index of the input voice. The configuration of the present invention will be described below based on an embodiment.

FIG. 8 is a block diagram for explaining one embodiment of the speech recognition apparatus according to the present invention. In the figure, 27 is a vector number sequence and class number sequence unit, and 28 is a vector distance vector matrix table classified by class. The same reference numerals as those in FIG. 7 are attached to the other portions having the same functions as those in FIG. As shown in FIG. 9, the standard pattern consists of a vector number sequence ▲ b ⁿ _j ▼ of the word n and a class number sequence ▲ c ⁿ _j ▼ of the word n into which each frame is classified. As shown in FIG. 10, the inter-vector distance matrix table 28 has a table of the number of classes M, and the element ▲ d ^m _ij ▼ of the matrix of the class m is when the input vector is the pseudophony i and the class m. Represents the distance between vectors when the vector of the standard pattern is the onomatopoeia j. DP
In the matching unit 26, the distance matrix table of the class corresponding to the standard pattern is selected from the class number sequence of 27 from 28, DP matching with the input pattern is performed with reference to the vector number sequence of the pattern, and the minimum distance is determined. The word identification section 29 determines the words that it has, and the recognition result output section 30 outputs them. That is, since the inter-vector distance matrix table in the recognition processing pre-calculates the inter-frame distance between the standard pattern and the quantized vector of the input voice, the vector number generation unit for the input vector during DP matching. A vector distance corresponding to the class number with the vector number and the class number for the standard pattern consisting of the vector number generated (converted) in 23 and the vector number sequence and the class number sequence classified into classes It operates so as to select a matrix table, derive the distance between vectors registered in the table, and search for the word (standard pattern) having the minimum distance.

Effect According to the present invention, therefore, the input speech is encoded by the vector number sequence, the inter-vector distance matrix table is pre-classified according to the type of phoneme, and the class-divided index (given by the class number sequence is given. The desired matrix table is selected based on the above (1) and DP matching is performed, so that voice recognition can be performed more quickly and accurately.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of a speech recognition apparatus, FIG. 2 is a diagram showing an example of speech analysis, FIG. 3 is a mapping model for time normalization, and FIG. 4 is time normalization by a distortion function. Figure 5, No.
FIG. 6 is a grid-like plan view for performing time normalization, FIG. 6 is a block diagram of a processor for performing DP matching processing, and FIG. 7 is a block diagram for explaining an example of the double split method. FIG. 8 is a diagram for explaining an embodiment of a voice recognition device according to the present invention, FIG. 9 is a diagram showing an example of the configuration of a standard pattern used for implementing the present invention, and FIG. It is a figure which shows an example of the distance matrix between vectors used for implementation. 20 ... Input part, 21 ... Analysis part, 22 ... Vector quantization part, 23 ... Template (vector) part, 24 ...
... Vector sequencer, 25 ... Vector distance matrix table, 26 ... DP matching section, 27
...... Vector number sequence and class number sequence part, 28 …… Class matrix-divided vector distance matrix table, 29 …… Word identification part, 30
...... Recognition result output section.

Claims (2)

[Claims] 1. A vector of a standard pattern is selected by quantizing an input voice, indexing the quantized input voice, selecting a vector-to-vector distance matrix table of a class corresponding to the standard pattern by a class number sequence of the standard pattern. A voice recognition method characterized in that DP matching with an input voice pattern is performed with reference to a number sequence. 2. Quantizing means for quantizing an input speech, indexing means for indexing the input speech quantized by the quantizing means, quantized vector number sequence and class numbers classified into classes. A standard pattern composed of a sequence, an inter-vector distance matrix table in which the inter-vector distance between the standard pattern and the input voice quantized by the quantizing means is pre-classified according to the type of phoneme, and the standard Speech recognition characterized by having DP matching means for selecting the inter-vector distance matrix table corresponding to the class number sequence attached to the pattern and matching the vector number sequence of the standard pattern with the index of the input voice. apparatus.