JPH0160160B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an improvement in a speech recognition device that recognizes monosyllables corresponding to Japanese kana characters etc. that are uttered in segments.

Conventionally, pattern matching has been used as a method for recognizing single syllables that are uttered separately. Typical methods include the following. Normally, most Japanese monosyllables are in the form of a consonant and a vowel, so a method is used to recognize them separately. First, the consonant and vowel parts of the speech pattern obtained by analyzing the input speech are extracted. A vowel category is determined by pattern matching the speech pattern of the vowel part, that is, the vowel pattern, with a pre-registered vowel standard pattern.

Next, a consonant category is determined by pattern matching the consonant pattern and a pre-registered consonant standard pattern, and a monosyllable recognition result is determined. Since stable patterns are generally obtained for vowel parts, the method of cutting out and the method of pattern matching do not pose much of a problem. On the other hand, there are various variations in the method of cutting out consonant patterns and the method of pattern matching.

For example, one possible method is to cut out a voice pattern of a certain number of frames from the beginning of the voice and use this as a consonant pattern. Since this method compares patterns with a fixed number of frames, there is no need to expand or contract the time axis. Therefore, recognition can be performed with a small amount of calculation.

However, the duration of consonants can vary greatly depending on the pronunciation or category. If the number of frames at this time is adjusted to the length of a long consonant, a large amount of memory will be required to store the standard pattern. Furthermore, in this method, when comparing consonants with short durations, the characteristics may be weakened because the comparison is performed using a long number of frames. On the other hand, if the number of frames is adjusted to the length of a consonant with a short duration, there is a drawback that it becomes difficult to recognize a consonant with a long duration.

A second method is to cut out the consonant part according to the length of the consonant, and match the consonant pattern and the standard consonant pattern by non-linearly expanding and contracting the time axis using a dynamic programming method. Using this method, it is possible to finely match patterns of different lengths.

However, this dynamic programming method requires a considerable amount of calculation. Also, unlike the case of word recognition, in the case of monosyllable recognition, the effect of nonlinear expansion and contraction of the time axis is thought to be small, so
This is not necessarily the best method.

An object of the present invention is to recognize monosyllables with a small amount of standard pattern memory and a small amount of calculation, and to obtain high recognition performance. In order to achieve this objective, the monosyllable recognition device according to the present invention includes an analysis unit that patterns monosyllables that are uttered separately and generates a speech pattern, and an analysis unit that extracts consonant parts and vowel parts from the speech pattern and creates a consonant pattern. and a standard pattern memory section that stores consonant patterns and vowel patterns of single syllables that have been uttered in advance as consonant standard patterns and vowel standard patterns, respectively; A vowel matching unit that matches vowel standard patterns to determine a vowel category; and a vowel matching unit that matches an input consonant pattern and the consonant standard pattern, and if the time lengths of the two are different, a vowel pattern is added after the shorter consonant pattern. and a consonant matching section that performs matching.

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the overall structure of an embodiment according to the present invention. The voice input from the microphone 1 is analyzed by the analysis section 2,
It is output as a voice pattern P. A consonant pattern C and a vowel pattern V are extracted from the voice pattern P in the voice extraction section 3. At the time of standard pattern registration, these patterns are held in the consonant standard pattern memory 4 and the vowel standard pattern memory 5, respectively. At the time of recognition, first the vowel pattern V is matched with the vowel standard pattern VR in the vowel matching section 6. This matching method only requires finding the distance between vowel patterns. As a result/
It recognizes vowel categories such as a/, /i/, u/, /e/, /o/, etc., and outputs the result as a vowel category VC. The consonant matching section 7 performs matching of consonant parts and outputs the recognition result as a consonant category CC such as /k/, /s/, etc. The operating principle of the consonant matching section 7 will be explained. FIG. 2 is a conceptual diagram for explaining an example of the operation of the consonant matching unit 7. Assume that the distance between the input consonant pattern 11 (the figure shows an example of a 5-frame pattern) and the consonant standard pattern 12 (the figure shows an example of an 8-frame pattern) is calculated. Each audio pattern is a time series of vectors with the time direction pointing to the right. From the start of the speech to the fifth frame, the input consonant pattern and the frames at the same position of the consonant standard pattern are compared as shown by arrow 13 to find the distance.
From the fifth frame onwards, the consonant standard pattern 12 is made to correspond to the vowel pattern 15 for one frame of the input voice as shown by the arrow 14, and the distance is determined. The rear part of the consonant pattern is usually almost continuous with the vowel pattern, and the vowel pattern is almost constant over time. Therefore, by making a vowel pattern follow the shorter consonant pattern, the original pattern can be reproduced.

This method allows patterns of different lengths to be easily matched. Further, since it is sufficient to separately store patterns for consonant parts and vowel parts as required as standard patterns, the amount of memory required for standard patterns can be reduced. The example shown here shows a case where the input consonant pattern is shorter than the standard consonant pattern, but in the opposite case, you can just replace both and follow the same procedure as above.

FIG. 3 is a block diagram showing a specific example of the configuration of the circuit of the consonant matching section 7. As shown in FIG. The consonant pattern C of the input voice is held in the consonant buffer 21 and the vowel pattern V is held in the vowel buffer 22. In addition, the consonant standard pattern CR is the consonant standard pattern buffer 23
The vowel standard pattern VR is held in the vowel standard pattern buffer 24. The input pattern frame counter 25 inputs the frame address IA to the consonant buffer 21.
Output. A consonant pattern CF for each frame is output from the consonant buffer 21 in accordance with this address. Similarly, the standard pattern frame counter 26 outputs a frame address RA to the consonant standard pattern buffer 23. In accordance with this address, the consonant standard pattern buffer 23 outputs the consonant standard pattern CRF for each frame. The input pattern frame counter 25 and the standard pattern frame counter 26 count up from 1 in synchronization, and when each reaches a value equal to the number of frames, the counter that has reached the same value stops counting. The data selectors 27 and 28 select the consonant pattern CF and the consonant standard pattern CRF while the frame counter continues counting, and output them to the distance calculation section 29. The distance calculation unit 29 calculates the distance between two patterns for each frame. The distance for each frame is the accumulator 3
When set to 0, the distance for one pattern is integrated.

A case where the number of frames of the consonant pattern is smaller than the standard consonant pattern will be explained. When the value of the input pattern frame counter 25 reaches the number of consonant pattern frames and stops counting, the data selector 2
7 selects the vowel frame data VF from the vowel pattern buffer 22 and outputs it to the distance calculation section 29. When the value of the standard pattern frame counter 26 reaches the number of frames of the consonant standard pattern, the distance between the two patterns has been determined, and the cumulative distance value d is output from the accumulator 30. The above is a case where the number of frames of the input consonant pattern is smaller, but the same operation is performed when the number of frames of the standard consonant pattern is smaller. A vowel standard pattern VRF is added behind the consonant standard pattern CRF. This distance integrated value d is input to the minimum value calculation unit 31 for each category of standard patterns. Here, the minimum value in all consonant categories is calculated, and the consonant category CC that has the minimum value is output as a recognition result.

The embodiments described above are merely examples selected for convenience of explanation, and the present invention is not limited to these embodiments.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a conceptual diagram for explaining an example of the operation of the consonant matching section, and FIG. 3 is an example of a specific circuit of the consonant matching section. FIG. In the figure, 1 is a microphone, 2 is an analysis section, 3 is a voice extraction section, 4 is a consonant standard pattern memory, and 5 is a consonant standard pattern memory.
is a vowel standard pattern memory, 6 is a vowel matching section, 7 is a consonant matching section, 11 is a consonant pattern,
12 is a standard consonant pattern, 15 is a vowel pattern, 21
22 is a consonant buffer, 22 is a vowel buffer, 23 is a consonant standard pattern buffer, 24 is a vowel standard pattern buffer, 25 is an input pattern frame counter, 2
6 is a standard pattern frame counter, 27 and 28 are data selectors, 29 is a distance calculation section, 30 is an accumulator, and 31 is a minimum value calculation section.

Claims (1)

[Claims] 1. An analysis unit that patterns the single syllables that are uttered in sections to create a voice pattern, a voice extraction unit that cuts out consonant parts and vowel parts from the voice pattern and creates consonant patterns and vowel patterns, and a monosyllable that has been uttered in advance. The consonant pattern and vowel pattern of
a standard pattern memory section that stores consonant standard patterns and vowel standard patterns, respectively; a vowel matching section that matches input vowel patterns with the vowel standard patterns and determines vowel categories; A single syllable recognition device comprising: a consonant matching section that adds a vowel pattern after the shorter consonant pattern and performs matching when the time lengths of the standard consonant patterns are different. .