JPH0830960B2 - High speed voice recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement of a high-speed voice recognition device for recognizing spoken voice at high speed.

(Prior Art) Conventionally, as a method for recognizing speech at high speed, for example, "High speed DP matching by beam search and vector quantization" is given on pages 33 to 44 of IEICE Technical Report SP87-26. Published in the following (hereinafter referred to as Reference 1)
Describes a method of performing DP matching with a small amount of calculation by using beam search.

This method is based on the recognition method based on pattern matching using DP matching. For example, if the distance between the frame i of the input pattern and the frame j of the standard pattern is d (i, j), the inter-pattern distance can be obtained by calculating the recurrence formula of DP matching shown below.

This recurrence formula is calculated in synchronization with the frame of the input voice pattern at time i. At this time, in order to reduce the amount of calculation, the value g (i, j) of the recurrence formula is calculated only in a probable portion by beam search. That is, the threshold θ (i) in frame i is And This is the beam width factor α added to the minimum value of g (i, j) in frame i. For g (i, j) larger than this threshold θ (i), frame i, Do not calculate as it is a correspondence of j. By such a method, the calculation of DP matching can be significantly reduced.

In addition to the DP matching described here, this beam search is performed by general pattern matching as described on page 142 of "Voice Recognition" (hereinafter referred to as Reference 2) published by Kyoritsu Shuppan Co., Ltd. Can be used for.

As one method for performing speech recognition by pattern matching using beam search on multiple processors in parallel, a set of words to be recognized (word set) is assigned to each processor and Then, there is a method of executing the same recognition processing in parallel on all the processors. According to this method, it is possible to execute the program at a higher speed than the execution by one processor without performing special execution management for the plurality of processors.

(Problems to be Solved by the Invention) When speech recognition by pattern matching using a beam search is executed in parallel by a plurality of processors, matching is performed with a certain input pattern in a word set assigned to one processor. If there are many words similar to the word that gives the minimum distance, the number of frames j of the standard pattern that give g (i, j) smaller than the threshold θ (i) will increase, and the beam width will not be sufficiently narrowed. For this reason,
There is a drawback in that the recognition time cannot be shortened because the load on one processor increases and the processor load is unevenly distributed, so that the plurality of processors cannot be used efficiently.

It is an object of the present invention to provide a high-performance high-speed voice recognition device that can be efficiently executed on a plurality of processors in voice recognition by pattern matching using beam search.

(Means for Solving the Problems) A continuous speech recognition apparatus according to the first invention of the present application, a speech analysis unit for analyzing an inputted speech to create a speech pattern, and a pattern using beam search for the speech pattern. Two or more recognition units that are recognized by matching, and
A standard pattern memory that holds a standard pattern for each of two or more recognition units, a result determination unit that outputs the best result among the recognition results from the two or more recognition units as an overall recognition result, and a standard pattern registration In some cases, a standard pattern distribution unit that distributes mutually similar patterns among the voice patterns to different standard pattern memories.

In addition to the first invention of the present application, a continuous voice recognition device according to the second invention of the present application includes a voice pattern memory for holding a plurality of voice patterns at the time of standard pattern registration, and a plurality of voice pattern memories in the voice pattern memory. It has a clustering unit that divides the voice pattern into clusters by clustering, and a standard pattern distribution unit that distributes the patterns in the one cluster to different standard pattern memories.

(Operation) The operation of the high-speed speech recognition device according to the present invention will be described. In the present invention, when the voice recognition processing by pattern matching using the beam search is executed in parallel on a plurality of processors, the recognition processing for the standard patterns of similar words can be executed on different processors. This increases the efficiency of beam search execution. The present invention is also effective for continuous word recognition using a phoneme as a recognition unit as shown in Document 2, but as an example, the present invention uses DP matching in units of words as described in Document 1 below. The method will be described below.

The high-speed speech recognition device according to the present invention comprises a plurality of recognition processors. The voice pattern obtained from the input voice is input to all recognition processors.
Each recognition processor recognizes the speech input by DP matching using beam search based on the standard pattern of the assigned word set. The recognition results from the respective recognition processors are compared, and the word having the smallest inter-pattern distance becomes the entire recognition result.

At this time, the processing efficiency varies depending on the method of assigning the word set to the recognition processor. According to the present invention, similar patterns among a plurality of standard patterns, that is, standard patterns having a small inter-pattern distance are distributed to different recognition processors. This prevents the number of frames of the standard pattern remaining in the beam from concentrating on one recognition processor, so that a plurality of processors can be used efficiently.

The first method for distributing the standard pattern is as follows. When registering a standard pattern, first, the input voice patterns are sequentially assigned to different recognition processors. When the standard pattern is assigned to all the recognition processors, the standard pattern to be registered next is assigned to the recognition processor having the largest distance among the pattern distances from the already assigned standard patterns to the recognition processor having the largest distance. As a result, one recognition processor can be assigned standard patterns having a large distance from each other.

In order to prevent the calculation amount and memory amount from being concentrated on one recognition processor, when the standard pattern assigned to a certain recognition processor exceeds a certain amount, the method of assigning the standard pattern to other recognition processors is recommended. Can be used.

As a second method of standard pattern distribution, there is a method of forming clusters of similar standard patterns by using a clustering method and distributing standard patterns included in one cluster to different recognition processors. As a clustering method, for example, "Pattern Information Processing" published by Corona
Various methods can be used, including the k-means method as shown on pages 113 to 120 (hereinafter referred to as Reference 3). As a result, standard patterns similar to each other included in one cluster can be assigned to different recognition processors. This method has a possibility of stable distribution compared with the first method, but since the clustering can be performed only after all the standard patterns are registered, the distribution is performed after all the standard patterns are registered. Becomes Also, a memory for storing the standard pattern is required.

(Embodiment) An embodiment of the high-speed speech recognition apparatus according to the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment according to the present invention. First, the operation at the time of recognition will be described.

The input voice is analyzed by the analysis unit 10 and is output as a time series of feature vector a (i); 1≤i≤I.
The feature vector a (i) is input to the three recognition processors 1 to 3. Although the number of recognition processors is three here, any number of two or more recognition processors can be used. Standard pattern memories 11 to 13 are connected to the recognition processors 1 to 3, respectively, and standard patterns to be recognized by the respective recognition processors.
bn; 1 ≦ n ≦ N is held. Recognition processors 1-3
The standard pattern memories 11 to 13 are the same for all three units.

In the recognition processors 1 to 3, when the feature vector a (i) is input, the DP matching calculation using the beam search as shown in Document 1 is executed. Feature vector bn (j) of frame j of the standard pattern read out from the standard pattern memories 11 to 13; a inputted as 1 ≦ j ≦ Jn
The distance d (i, j) between (i) is calculated, and the recurrence formula of the formula (1) is calculated based on this. At this time, the threshold value θ is read from the threshold value calculation unit 4, and the beam search is performed using this value as the threshold value. That is, the value g of the recurrence formula is the threshold value θ.
In the above cases, recurrence calculation is not performed. In the threshold calculation unit 4, the minimum value of the recurrence formula value g in the frame i−1, which is one frame before the present, in each of the recognition processors 1 to 3
gmin (k); 1 ≦ k ≦ 3 is read out, and the threshold value θ is obtained from the minimum value in each recognition processor by the following equation.

These processes are the feature vector a of the input voice.
The value gn of the recurrence formula in the final frame Jn of the standard pattern left in the beam, which is performed up to frame I in (i)
Then, the word name W that gives it is output to the result determination unit 5. In the result determination unit 5, the word R giving the minimum value in the values gn of these recurrence formulas is output as the recognition result.

Next, the operation when registering the standard pattern will be described. The speech of the word set generated during registration is analyzed by the analysis unit 10
The voice pattern obtained by the analysis is distributed to any of the standard pattern memories 11 to 13 by the standard pattern distribution unit 6 and held as the standard pattern b.

FIG. 2 shows the standard pattern distributor 6 according to the first invention of the present application.
FIG. Standard pattern b for N types of words
A case of recognizing n; 1≤n≤N using three recognition processors will be described. The distribution unit 61 sequentially outputs the first three standard patterns bk; 1 ≦ k ≦ 3 to the recognition pattern memories 11 to 13. When a new standard pattern to be registered next is input, the distance calculation unit 62 sequentially calculates the inter-pattern distances with the standard patterns in the already-assigned standard pattern memories 11 to 13 and The minimum value for the memory is obtained and output as the minimum value Dmin (k) for the standard pattern memory of number k. Then, the maximum value calculation unit 63 obtains the standard pattern memory number k at which the minimum value Dmin (k) is maximum. Distributor 61
Then, the new standard pattern is distributed to any of the standard pattern memories 11 to 13 according to this k.

FIG. 3 shows the standard pattern distributor 6 according to the second invention of the present application.
FIG. First, the voice pattern obtained by analyzing the voice uttered for registering the standard pattern is held in the voice pattern memory 64. When the voice patterns of all the word sets are registered, the clustering unit 65 classifies the N standard patterns held in the voice pattern memory 64 into M clusters by the clustering method. The clustering method used here is the k-means method as shown in Reference 3. M is an arbitrary number equal to or less than N. Then, the distribution unit 66 sequentially distributes the standard patterns classified into one cluster into the standard pattern memories 11 to 13. That is, Nm standard patterns bm (n) classified into cluster m; 1 ≦ n ≦ Nm are standard patterns bm (1)
Is a standard pattern memory 11, bm (2) is a standard pattern memory 12, and so on. By distributing in this way, similar standard patterns can be distributed to different standard pattern memories.

(Effects of the Invention) According to the present invention, when performing recognition processing using beam search on a plurality of recognition processors, a high-performance high-speed speech recognition apparatus can be realized by increasing the efficiency of beam search. .

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG.
FIG. 3 is a configuration diagram showing an embodiment of a standard pattern distribution unit according to the first invention of the present application, and FIG. 3 is a configuration diagram showing an embodiment of a standard pattern distribution unit according to the second invention of the present application. In the figure, 1,2,3, ... Recognition processor, 4 ... Threshold value calculation unit, 5 ... Result determination unit, 6 ... Standard pattern distribution unit, 10 ... Analysis unit, 11,12,13
... standard pattern distribution unit, 61, 66 ... distribution unit, 62 ... distance calculation unit, 63 ... maximum calculation unit, 64 ... voice pattern memory, 65 ... clustering unit, respectively.

Claims (2)

[Claims] 1. A voice analysis unit for analyzing a voice input to create a voice pattern, two or more recognition units for recognizing the voice pattern by pattern matching using a beam search, and the two or more recognitions. A standard pattern memory that holds a standard pattern for each set, a result determination unit that outputs the best result among the recognition results from the two or more recognition units as an overall recognition result, and a voice when the standard pattern is registered. A high-speed speech recognition apparatus comprising: a standard pattern distributor that distributes patterns that are similar to each other to different standard pattern memories. 2. A voice pattern memory for holding a plurality of voice patterns at the time of registering a standard pattern, a clustering unit for dividing a plurality of voice patterns in the voice pattern memory into clusters by clustering, and one of the one clusters. 2. The high-speed speech recognition apparatus according to claim 1, further comprising a standard pattern distribution unit that distributes each pattern of FIG.