JPS6172297A - Standard pattern generation system for voice recognition - 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 Technical Field The present invention is particularly useful for continuous speech recognition of a specific speaker using a pattern matching method, and is particularly useful for recognizing continuous speech of a specific speaker using a pattern matching method.

A method of creating multiple standard patterns for each category by taking into account variations in patterns due to utterances.In speech recognition methods for non-specific speakers, conventional speech recognition methods have traditionally used clustering analysis from the speech patterns of a large number of people. A technique is used in which several clusters are determined and a pattern representing each cluster is used as a standard pattern. In this case, since clustering analysis is performed, the target is a case where there are a large number of voice patterns.

On the other hand, the problem of the present invention is mainly the creation of standard patterns for a specific speaker, and the method of selecting a plurality of standard patterns from a relatively small number of registered speech patterns. As a result, there were problems with clustering analysis, and the algorithm was too complex to incorporate into a recognition device.

<Objective of the Invention> In speech recognition of a specific speaker, it is desirable to have a plurality of standard patterns for each category, taking into account variations in pronunciation and the effects of articulatory combination. - As an example, let's consider the case of recognizing monosyllables/CV/units.

At this time, for example, the registered speech category monosyllable / k a
When creating a standard pattern for /, in addition to isolated utterances, /Vka/(V=a, i, u, e.

0) etc. and register them, a problem arises in which, for example, it is desired to finally register three representative patterns as standard patterns.

The present invention is a method for selecting a designated number of representative voice patterns from a relatively small number of voice patterns in a certain registered voice category, and is particularly capable of selecting a specified number of representative voice patterns based on a certain quantitative judgment logic. The present invention provides a useful method that can be effectively used in a speaker-specific standard device having a voice registration function.

<Example> The present invention will be explained in detail below based on examples.

A speech pattern is represented by a time series of feature parameters, and there is a distance d ((, y
) is given.

Possible feature parameters include autocorrelation coefficients and cepstral coefficients, and distances can be calculated using dynamic programming using Euclidean distance, but the following explanation is based on generality that does not depend on these. It is.

In order to facilitate understanding of Arborim, the following explanation will include the case where one standard pattern is used for reference.

l) When using one standard pattern Figure 3 shows its conceptual diagram.

A certain category C voice pattern (Den 12% 2゜...
r%N). Of these, there are 71 standard putters.
The criterion for selecting is that for each electric current, D = Σd(, j
, Electric 1) Find 3=+, and select the voice pattern Xi that gives the minimum Dl among (1), (j=+,...,N)) for category C.
This is the standard pattern.

This is based on the idea of the least squares principle, and the standard pattern X is selected as being close to the "center of gravity" of the speech pattern.

The above is the basic algorithm, but when setting a plurality of standard patterns as in the present invention, the algorithm is as follows.

2) When there are two or more standard patterns From the voice pattern of category C (l,...,xHI, there are generally m standard patterns "1 (11,...).

The case where 'i(ml') is selected will be described. In Figure 1, m=2
A conceptual diagram of the case is shown.

m standard pattern candidates from (den,...,den,),
Take out R, - (whisper, , electric, ) while 1”' i
(ml N'.

A set R of speech pattern l and m standard pattern candidates
1=(expletive 1(11'..., 2%i(ml''s "distance" D
(x, ,R,), D (x, R,)=m' n [d(qj+x
1 (kl), k=1. −, ml” 1 (min represents the minimum value). Then, using this minimum distance
Set of standard pattern candidates R=(・i(+1'...
・,・1−)), D,= Σ D (x J
Find R1=(%i(+
1'..."i(ml') is selected as m standard patterns.

In this method, as mentioned above, if the arrangement (α(X-, ), the desired combination of standard patterns can be obtained.

FIG. 2 shows an example of the configuration of a speech recognition device according to this method.

This speech recognition device has a registration mode and a recognition mode.

In the registration mode, input audio data is sent to microphone 1. The signal passes through an amplifier 2 and an A/D converter 3, is converted into a parameter series by a feature parameter extraction section 4, and is stored in a speech pattern memory 5. If we decide to register each category, the voice pattern memory 5 will be
It only needs to be large enough to accommodate the audio pattern of a certain category.

The mutual distances of the patterns in the voice pattern memory 5 are determined by the matching calculation unit 6 and stored in the distance matrix memory 7. The logic judgment unit Cll18 is
Generate a combination R of standard pattern candidates, calculate the above-mentioned Dl by adding the minimum value calculation from the distance matrix memory 7, determine the combination with the minimum value, and voice the audio pattern of that combination. It is transferred from the pattern memory 5 to the standard pattern memory 9.

In the recognition mode, the input voice is converted into a time series of parameters by the feature parameter extraction unit 4, the distance from the pattern in the memory 9 for the prenatal pattern is determined by the matching calculation unit 6, and the input voice is recognized by the logic determination unit [1110]. Judge the results.

<Effects of the Invention> As described above, according to the present invention, it is possible to create a predetermined number of standard patterns from a relatively small number of speech patterns, and it is also possible to create this based on a certain quantitative judgment logic.
In particular, a very useful method can be provided in a specific speaker recognition device.

[Brief explanation of drawings]

Figure 1 is a conceptual diagram showing an example of the present invention when there are two standard patterns, Figure 2 is a professional diagram showing an example of the configuration of a speech recognition device, and Figure 3 is a standard pattern for reference. It is a conceptual diagram when there is one pattern. C...Category, X + X2..., xn
・Voice pattern!

Claims (1)

[Claims] 1. In a speech recognition method using pattern matching, which creates a specified number of standard patterns for each registered voice category, it is possible to create a specified number of standard patterns for any combination of registered voice patterns in each category as a standard pattern. An evaluation value indicating appropriateness is obtained by summing the minimum distance between one voice pattern in the category and the voice pattern in the combination over all voice patterns in the category, and the voice pattern has the minimum evaluation value. A standard pattern creation method for speech recognition that uses a combination of as the standard pattern for that category.