JP3094473B2 - Dynamic programming collator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a dynamic programming method for associating a standard pattern with an input pattern.
amming (hereinafter referred to as DP).

[0002]

2. Description of the Related Art Speech recognition aims at observing continuously generated speech signals and converting them into discrete language symbols of phonemes and syllables. The sound pattern is characterized by being one-dimensional along the time axis as compared with the character or image pattern. Moreover, even if the same person utters one word, none of the words has the same length but changes in time length.

As a word speech recognition method, a method of recognizing a word by matching a word standard pattern with an input speech pattern is often used. As a standard pattern,
It is prepared in advance in the form of a time series of a spectrum obtained by processing the input. The similarity with the input utterance data is calculated, and the one with the highest similarity is set as the uttered word.

[0004] By the way, even if the same word is used, the time length often differs each time it is uttered. Although the time length of consonants is relatively stable, the time length of vowels varies greatly. Therefore, nonlinear pattern matching on the time axis is required. In the matching that normalizes the time axis, a standard pattern is prepared as a time series of feature amount parameters, and the non-linear correspondence with the input pattern is examined to absorb the fluctuation of phonemes on the time axis due to utterance It is. In this case, the correspondence between the standard pattern and the input pattern on the time axis is examined one by one. Dynamic Programming (DP)
Is used. This DP has been described in many books (for example, pattern recognition, supervised by Kenichi Mori, The Institute of Electronics, Information and Communication Engineers, Corona Co., Ltd., first edition issued on November 1, 1988), and the description thereof is omitted.

A specific example of such a DP comparison will be described with reference to the drawings. FIG. 11 is a block diagram showing a configuration of a conventionally used DP matching apparatus, and FIG. 12 is a detailed view of the DP matching unit 15 shown in FIG. FIG. 13 is a flowchart showing the operation of the DP matching device. Description will be made focusing on the operation of FIG. First, an input time-series pattern is read (step 20). Thus, the input time-series pattern read by the input time-series pattern reading unit 10 is stored in the input pattern memory 11. Next, one standard time series pattern is read (step 21). Many standard time series patterns are stored in the standard time series pattern storage unit 12, and one of them is read out by the standard time series pattern reading unit 13 and stored in the standard pattern memory 14. Next, DP collation is performed (step 22).

[0006] To explain the DP collation, the DP shown in FIG.
A description will be given of a path calculation as an example. In FIG. 14, the local distance d
(I, j) is the distance between the i-th feature of the input time-series pattern and the j-th feature of the standard time-series pattern. As the distance, a Eugrid distance or a city distance between these feature values is used. Cumulative distance G (i, j)
Means the local distance d of the grid point among various routes starting from the base point and reaching the grid point (i, j) shown in FIG.
Shows the path where the sum of (i, j) is the minimum. (B) 3
The two initial conditions are based on the calculation of the cumulative distance G (i, j) as the base point (0
, 0), and the formula for calculating G (i, j) is as follows: G (i, j) is the local distance d (i, j) of grid point (i, j) and FIG. ), G (i)
-1, j) and the smaller value of G (i-1, j-1).

[0007] A DP path is formed by connecting the base point given as the initial condition the cumulative distance G (i, j) as a start, and FIG. 15 shows a standard time series pattern 1 to an input time series pattern. The result of collation with n is shown. As shown in FIG. 7A, the matching operation is performed for each feature amount (0 to n) of the standard time-series pattern and the local distances d (i, 0), d (i,
1)... D (i, n) is obtained, and then the cumulative distance G (i, n) is obtained.
0), G (i, 1)... G (i, n) are obtained. Using these, a DP path from the base point (0,0) is obtained. The processing unit in FIG. 7A indicates this collation work unit.

FIG. 12 is a detailed diagram of the DP collating unit 15, and the input pattern feature amount i read from the input pattern memory 11 is shown.
The local distance d (i, j) between the standard pattern feature amount j read from the standard pattern memory 14 and the local distance calculation unit 151
Is calculated. The cumulative distance memory 154 is composed of a vector register, and G (i−1,0) to G (i−1,0) calculated so far.
The cumulative distance to G (i-1, n) is stored. The minimum cumulative distance calculation unit 153 calculates G (i−1, j−1) and G
The smaller value of (i−1, j) is taken out, and this value and the local distance d (i, j) are added by the adder 152 to obtain G
(I, j) is obtained. This G (i, j) is stored in the cumulative distance memory 154. The control unit 155 controls these operations.

FIG. 16 is a timing chart showing the operation of the DP matching unit 15 of FIG. The DP collating unit is composed of a vector processor, and performs a pipeline process on the objects to be processed on the vector register 154 collectively. The processing unit shown in FIG. 16 is a processing unit shown in FIG. 15 (a) and represents a processing operation on the characteristic amount (0 to j to n) of one standard time-series pattern with respect to the characteristic amount i of the input time-series pattern. The overhead means that the reading of all G (i-1, j-1) of the feature amount i of the input time-series pattern in the processing unit 1 is completed, and the first G (i, j-) of the feature amount i + 1 in the processing unit 2 is completed.
This is a period until the reading of 1) is started. In this way, all the standard time-series patterns are collated (step 23).

[0010]

As described above, in the collation processing, the objects to be processed in the vector register 154 are collectively pipelined by the vector processor, but the processing is performed for each standard time series pattern as a processing unit. I have. In the case of the normal vector register 154, there is a capacity to store about several hundred standard time-series patterns. However, if processing is performed in units of one standard time-series pattern, the function of the vector register is not sufficiently used. Processing efficiency is poor. Vector registers 154 for each processing unit
, The number of times of loading increases. Further FIG.
As shown in (1), the overhead time increases for each processing unit.
For this reason, even if a vector processor is used, the advantage cannot often be fully exploited.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned problems, and uses a vector processor capable of performing batch processing for DP matching processing, and reads a plurality of standard time-series patterns into a vector register and performs batch vector matching processing. It is an object of the present invention to provide a dynamic programming collation apparatus which improves the processing efficiency.

[0012]

FIG. 1 is a diagram illustrating the principle of the present invention. In FIG. 1, reference numeral 1 denotes an input pattern reading unit which reads a feature amount of an input pattern to be processed from an input time-series pattern. Reference numeral 2 reads a plurality of standard time-series patterns for performing matching with respect to a feature amount of an input pattern to be processed by a plurality of standard pattern reading units. Reference numeral 3 denotes a collation unit that collates a plurality of standard time-series patterns with respect to the feature amount of the input pattern.

The standard time-series pattern has a network structure. In addition, the matching with the input time-series feature amount for the common network portion of the network structure is performed sequentially and continuously.

A plurality of storage units are provided in the plurality of standard pattern reading units 2 to store the comparison result of the comparison unit 3, and the comparison unit 3
Can simultaneously read the contents of the plurality of storage units.

[0015]

The matching unit 3 sequentially performs the matching process on the feature amounts of a plurality of standard patterns with respect to the feature amount of the input pattern, thereby reducing the number of times of reading the standard time-series pattern and the overhead, and improving the matching process efficiency. improves.

Further, by forming the standard time series pattern into a network structure, the parts constituting the network can be processed in the same processing unit, thereby improving the processing efficiency. Also,
The processing efficiency is improved by collating the common network portion with the feature amount of the input pattern in the same processing unit.

A plurality of storage units are provided in the plurality of standard pattern reading units 2, and the same comparison result of the comparison unit (3) is stored in the plurality of storage units, so that a plurality of accumulated distances, for example, G (i −1, j−1) and G (i−1, j) can be read simultaneously, so that the processing speed is improved.

[0018]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 2 is a block diagram showing the configuration of the first embodiment of the present invention. Reference numeral 10 denotes an input time-series pattern reading unit that reads a human utterance. Reference numeral 11 denotes an input pattern memory which reads out and stores a feature amount of an input time-series pattern to be subjected to a matching process from an input time-series pattern reading unit 10. Reference numeral 12 denotes a standard time series pattern storage unit 12, which stores a large number of standard time series patterns such as words. Reference numeral 16 denotes a plurality of standard time-series pattern reading units, which simultaneously read a plurality of standard time-series patterns to be compared with the input time-series patterns into the standard time-series patterns from the storage unit 12. Reference numeral 17 denotes a plurality of standard time-series patterns, which stores a plurality of standard time-series patterns read by the plurality of standard time-series patterns reading unit 16. Reference numeral 18 denotes a plurality of DP matching units which sequentially perform DP matching of a plurality of standard time-series patterns read from the plurality of standard pattern memories 17 with respect to the feature amount of the input time-series pattern read from the input pattern memory 11. In addition, this plural D
The detailed structure of the P collation unit 18 is basically the same as the configuration shown in FIG. 12, and the result of collation with a plurality of standard time-series patterns is stored in the cumulative distance memory 154.

FIG. 3 is an operation flowchart of this embodiment. First, the input time series pattern is read in the input pattern memory.
11 (Step 30), and then read the plurality of standard time-series patterns into the plurality of standard pattern memories 17.
(Step 31). For each feature of the input time-series pattern, first match with the feature of the first standard time-series pattern, and then compare with the feature of the next standard time-series pattern. Check with the sequence pattern.

FIG. 4 is a diagram showing how the input time series pattern is compared with a plurality of standard time series patterns. FIG. 15 shows a processing operation for performing a matching process for each standard time-series pattern. FIG. 4 shows a state in which the plurality of standard time-series patterns are sequentially processed in one processing unit.

In the case of this embodiment, the processing conditions shown in FIG. 14 are as follows. This is set so that the DP path of each standard time series pattern starts from the base point (0, 0). G (0, j) = 0 (j = boundary of each standard time series pattern) = {(j} boundary of each standard time series pattern) Cumulative distance; G (i, j) = d (i, j) + min ｛G (i-1, j), G
(I-1, j-1)｝ (j 境界 boundary of each standard time-series pattern) G (i, j) = ∞ (j = boundary of each standard time-series pattern) In this way, the collation for all standard patterns is completed. The collation processing is performed until (step 33).

FIG. 10 is a timing chart of G (i-1,
j-1) and G (i-1, j) are read at the same time. If they are read by shifting one clock as shown in FIG. 16, the timing chart of this embodiment is obtained. . When comparing FIG. 10 with FIG. 16, in FIG. 10, since the standard time-series patterns 1 to n are compared with the input pattern feature amount i during the processing unit 1, the overhead is between the next processing unit 2. , And the reading of the standard time-series patterns 1 to n is also performed collectively. On the other hand, the overhead of the collation processing of the conventional method shown in FIG. 16 is generated for each collation with each standard time series pattern, and the standard time series patterns 1 to n are also read, so that the number of times of reading is n. For this reason, as in the present embodiment, the comparison between a plurality of standard time-series patterns and the feature amount of the input time-series pattern is performed within one processing unit, thereby reducing the overhead and the number of times of reading the standard time-series pattern, and performing a matching process. Processing efficiency is improved.

FIG. 5 shows a case where the standard time series pattern has a network structure. Movies and heroes, and English is pronounced both as a and a. Males are also pronounced as Yu and Yu. In such a case, by collating the rays A and A together, the efficiency of the collation processing is improved. FIG. 6 shows the state of the matching process.

FIG. 7 shows a state in which a common part of a standard pattern having a network structure is shared. The movie ray (A) and the hero ray (A) are common to the painting and the male. FIG. 8 shows how the common part is subjected to the collation processing in common. The processing amount is reduced by performing the common processing in this manner.

FIG. 9 is a block diagram showing a detailed configuration of the multiple DP collating unit 18 according to the second embodiment of the present invention. Other configurations are the same as those of the first embodiment. The difference from FIG. 12 is that two memories having the same capacity are provided as the cumulative distance memories 184 and 185,
By storing the same contents, G (i-
1, (j), G (i-1, j-1) can be read out at the same time, thereby speeding up the processing.

FIG. 10 is a timing chart of the multiple DP collating unit 18 shown in FIG. The difference from FIG. 16 is that the matching is performed with the standard time-series patterns 1 to n in the processing unit 1, and the overhead is generated only for each of the processing units 1, 2,.
Number of standard time series patterns per processing unit compared to 16 n-
The overhead time is reduced by one, and G (i
−1, j−1) are simultaneously performed, so that the matching process is speeded up by one clock per one standard time-series pattern.

[0027]

As is apparent from the above description, the present invention performs the matching process between the feature quantity of the input time-series pattern and a plurality of standard time-series patterns in one processing unit.
The collation processing efficiency can be improved.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a block diagram illustrating a configuration of a first embodiment.

FIG. 3 is a flowchart showing the operation of the first embodiment.

FIG. 4 is a diagram illustrating a collation process according to the first embodiment.

FIG. 5 is a diagram illustrating a standard pattern having a network structure.

FIG. 6 is a diagram illustrating a process of collating a standard pattern having a network structure.

FIG. 7 is a diagram illustrating the sharing of a standard pattern having a network structure.

FIG. 8 is a diagram illustrating a collation process in which the standard patterns shown in FIG. 7 are shared.

FIG. 9 is a block diagram illustrating a configuration of a multiple DP matching unit according to a second embodiment.

FIG. 10 is a timing chart of the multiple DP matching unit of the second embodiment.

FIG. 11 is a block diagram illustrating a configuration of a conventional DP matching device.

FIG. 12 is a diagram illustrating a detailed configuration of a DP matching unit in FIG. 11;

13 is a flowchart showing the operation of the device shown in FIG.

FIG. 14 is an explanatory diagram of DP path calculation.

15 is a diagram illustrating a method of performing a matching process by the device illustrated in FIG. 11 and a matching result.

16 is a timing chart of the DP matching unit shown in FIG.

[Explanation of symbols]

 10 Input time series pattern reading section 11 Input pattern memory 12 Standard time series pattern storage section 16 Multiple standard time series pattern reading section 17 Multiple standard pattern memories 18 Multiple DP matching section

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-1-144100 (JP, A) JP-A-2-250188 (JP, A) JP-A-63-226993 (JP, A) (58) Survey Field (Int.Cl. ⁷ , DB name) G10L 11/00-21/06 JICST file (JOIS)

Claims (4)

(57) [Claims] A dynamic programming matching apparatus for matching a standard time series pattern with an input time series pattern, wherein a plurality of standard pattern reading units for reading a plurality of standard time series patterns.
When an input pattern reading unit for reading the input time series, and feature quantity of the read input time series of the input pattern read unit
The matching with the feature values of the plurality of standard time-series patterns is input.
Force time series components and a characteristic amount for each sequentially performed continuously, dynamic programming matching system characterized by comprising a verification portion formed of a single vector processor. 2. The dynamic programming matching apparatus according to claim 1, wherein said standard time series pattern has a network structure. 3. The dynamic programming matching apparatus according to claim 2, wherein the matching of the standard time series pattern with the feature amount for a common network portion is performed in common. Wherein said plurality standard pattern reading unit has a plurality of storage portions storing the collation result of the collating unit, the comparison unit
2. The dynamic programming collation apparatus according to claim 1, wherein the contents are simultaneously read from the plurality of storage units.