JPS6193499A - Voice pattern collation system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voice pattern matching method in a voice recognition device.

Conventionally, in speech recognition using halting pattern matching, the speech to be recognized is registered in advance, the input speech and the registered speech patterns are compared and matched to find the most similar speech, and the recognition results are calculated. It is said that In this case, when a voice is input, it takes time to match all the registered sounds, so a method is known in which it is determined which registered pattern should be matched based on other characteristics and then the pattern is matched.

As a feature for this judgment, there is a method that uses whether or not there is a silent part in the voice section (Japanese Patent Laid-Open No. 513-14
No. 2598), speech that includes silent parts is compared only with registered patterns that have silent parts, thereby reducing the number of targets and achieving high-speed recognition. This takes advantage of the fact that there is a silent interval before the / sound or consonant in /p/, /l/, /.
In the case of word speech recognition, since many words tend to have silent sections, there is a disadvantage that matching takes longer when the word is classified as having a silent section than when it does not.

Furthermore, when using the presence or absence of a silent section, there is a drawback that the number of targets cannot be reduced correctly because the occurrence of silent sections before the /d/ and /g/ phonemes is not constant.

The present invention was made in order to solve the above-mentioned drawbacks of the prior art, and in particular, the present invention was made with the aim of correctly distinguishing words that have silent sections in speech and recognizing them at high speed. be.

In order to achieve the above object, the present invention provides a means for converting an audio signal into an electrical signal, a means for converting the electrical signal into a characteristic parameter, and a means for registering the audio pattern created in this way. In a speech recognition device that outputs a recognition result by comparing and matching an input speech pattern with a pattern that has already been registered, the number of characteristic parts of the pattern is added to a part of the pattern and registered. The combination of information added from multiple patterns for one voice is registered together with the pattern, and the degree to which the number of the same characteristic parts matches when an unknown input pattern is given is registered. The feature is that the pattern is compared with the additional information, and only the standard patterns that match to a certain degree or more are matched. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention, in which 1 is a microphone, 2 is a feature extraction section, 3 is a voice section detection section, 4 is a power detection section, 5 is a register, 6 is a certain value or less determination unit, 7 is a counter, 8 is a silent number pattern, 9 is a first matching unit, 10 is a word dictionary, 11 is a certain value or less determination unit, 12 is a second matching unit, 13 is the maximum similarity The calculation unit 14 is a result output unit. Below, a case will be explained in which a silent section is adopted as a characteristic part of the pattern. However, the present invention is not limited to silent sections, and there are many high frequency components. You may pay attention to any features such as the voiced portion, the voiced portion, or the unvoiced portion.

First, to explain dictionary creation, it is known that in speech recognition for unspecified speakers, one dictionary pattern is registered by averaging the voices of several speakers (Speech Study Group Materials 581-57). The number of silent sections is added to each pattern as shown in FIG. Figure 2 (a) shows the number of silent intervals of a certain word by speaker A, and the 8-bit memory M represents the number of silent intervals 0, 1, 2, . . . 7 in order, and (
a) indicates that there was one silent section. An 8-bit pattern like this is attached to a part of the audio pattern. Speakers B and C use the same word, but the number of silent intervals is O.
Met. Here, the average of the patterns of eight people is used as a dictionary pattern, and when this 8-bit pattern is summed together with the pattern, a dictionary pattern as shown in (i) is created. Each element is represented by 3 bits, and 6 out of 8 people said the silent section was 0.2 people were 1, and no one else had more. The cross section of (i) is made continuous and shown in (j). Such patterns are registered together with dictionary patterns.

Returning to Figure 1, the features of the voice are extracted using a group of bandpass filters, etc., and the voice section is stored in a register.
By detecting the power of the voice, etc., we use a counter to count the number of silent sections that are below a certain value, as shown in Figure 2 (k).
Create a silent pattern like this. The above-mentioned constant value is not necessarily 0, but may be determined to be a value that can be considered as a silent section, taking into consideration surrounding noise and the like. Next, figure 2 of the dictionary (i
) and the input part (k) are matched by the first matching unit. This matching only requires calculating the sum of products of corresponding elements. If this value is less than a certain value, for example O, there were no 291 putters with the same number of silent intervals, so there is no need to match the entire pattern, and it is determined that the pattern is different and the next dictionary pattern is used. Take out. The result of this matching is the second one only for dictionary patterns that have a certain value.
The matching section matches the entire pattern. In this way, all the patterns registered in the dictionary are compared, and the one with the highest degree of similarity is taken as the recognition result.

On the other hand, as shown in FIG. 3, a multiplication circuit 15 is provided, and instead of pre-selectively using the number of silent intervals, a new similarity is determined by the product of the similarities obtained in the first and second matching. You may do so. In this case, if either of the matching results is 0, there is no need to take the product, or it is possible to decide on a dark value of similarity, and if it is greater than that value, then abort the subsequent calculations and use it as the recognition result.

FIG. 4 is a block diagram for explaining another embodiment of the present invention, in which 16 is a section end determination section, 17 is a preliminary selection section,
Reference numeral 18 is a collation section, and other parts having the same functions as those in FIGS. 1 and 3 are given the same reference numerals as in FIGS. 1 and 3. In FIG. 4, the word dictionary records the patterns of each registered word as well as the range of change in the number of silent intervals in those words. The range of this change is the variation in the number of silent intervals when the user utters the same word several times in the case of a device suitable for specific speakers, and the variation in the number of silent intervals when the user utters the same word several times, and the variation in the number of silent intervals when the user utters the same word several times in the case of a device suitable for non-specific speakers. The degree of variation in the number of silent sections is registered in advance.

For example, the word ``upper'' is memorized as 0.0 because there is no silence no matter who utters it, and ``lower J/5ita/ has a silent part before /1/, so it is recorded as 0.0 regardless of the person who utters it. does not change, and in this case 1.1 is memorized.On the other hand, "Right J
In /m1g1/, some people have a silent section before /g/ and some do not, and since 0 or 1 can occur, 0 and 1 are stored. Words that vary from one to three depending on the person: 1.
Remember it as 3.

First, when an unknown voice is input from the microphone shown in FIG. 1, features are extracted using a means such as a group of band-pass filters, and only the section of the voice is transmitted to the next step. Next, in the case of a group of band-pass filters, if there is a portion where the total output of all channels, that is, the power, is less than a certain value, the counter is incremented by one. It counts how many parts of the entire word are below a certain value, and the counter does not advance even if there are consecutive parts below a certain value. This constant value can be determined by determining the silent section detection level based on surrounding noise, etc. In this way, find how many silent intervals there are in the whole word,
Find and match the silent interval in the word dictionary. For example, if the number of silent intervals in the input word is "2", 0,2
; 1, 2; 2, 2; 2.3; 1, 3, etc., are matched against words that include 2 within the variation range, and word patterns that do not include 2 are not matched. The maximum similarity obtained through matching is output as the recognition result.

The matching method in this case is not limited; well-known DP matching (matching using dynamic programming) or the like may be used.

Processing As is clear from the above description, according to the present invention, the number of characteristic parts is statistically processed, thereby eliminating the possibility of selecting a wrong target.

[Brief explanation of the drawing]

FIG. 1 is an electrical block diagram for explaining an embodiment of the present invention, FIG. 2 is a diagram for explaining dictionary registration operation, and FIGS. 3 and 4 are respectively diagrams for explaining an embodiment of the present invention. FIG. 7 is a configuration diagram for explaining another embodiment. DESCRIPTION OF SYMBOLS 1... Microphone, 2... Feature extraction part, 3... Voice section detection part, 4... Power detection part, 5... Register,
6... Below a certain value determination unit, 7... Counter, 8...
・Silent number pattern, 9...first matching section, 10...
Word dictionary, 11... Constant value or less determination unit, 12... Second matching unit, 13... Maximum similarity calculation unit, 14... Result output unit, 15... Multiplication circuit, 16... - Section end determination section, 17...preliminary selection section, 18... collation section. Figure 1 (C) Lokao Tech 4 Old C Ck) Mouth Capable W Old E Figure 3 Figure 4

Claims (3)

[Claims] (1) The system includes a means for converting an audio signal into an electrical signal, a means for converting the electrical signal into a characteristic parameter, and a means for registering the audio pattern created in this way, and the input audio pattern and In a speech recognition device that compares and matches patterns that have already been registered and outputs recognition results, the number of characteristic parts of the pattern is added to a part of the pattern, and multiple patterns are combined for one speech to be registered. The combination of added information is registered together with the pattern, and when an unknown input pattern is given, the degree to which the number of the same characteristic parts matches is compared with the added information of the registered pattern, A voice pattern matching method characterized by matching only standard patterns that match to a certain degree or more. (2) When an unknown input pattern is given, the degree to which the number of the same characteristic parts matches is compared with the additional information of the registered pattern, and the degree of matching is reflected in the similarity. A voice pattern matching method according to claim (1). (3) The change range of the number of characteristic parts is added to a part of the pattern, and when an unknown input pattern is given, the number of the same characteristic parts falls within the change range of the standard pattern. The voice pattern matching method according to claim 1, characterized in that only those that are present are matched.