JPH06149286A - Unspecified speaker speech recognizing device - Google Patents

Abstract

PURPOSE:To provide the unspecified speaker speech recognizing device which has high robust performance to variation on the time base and frequency base of a specific word and also has a high recognition rate. CONSTITUTION:This speech recognizing device is equipped with speech analyzing means (11-14) which analyzes inputted speech signals and generate monitor waveform data, and a data recognizing means 15 having a means which generates one identification system identified with data for identification corresponding to the specific word that one unspecified person voices and plural evaluation systems identified with data for evaluation corresponding to specific words that other speakers voice, and a means which finds the minimum sum of squares of the differences between similarity estimated values obtained by inputting respective evaluation data to corresponding evaluation systems for the specific words and similarity estimated values obtained by inputting the respective evaluation data to the identification system.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an unspecified speaker voice recognition device for recognizing the sound of a specific word emitted by an unspecified speaker.

[0002]

2. Description of the Related Art In a conventional speaker-independent voice recognition device, in order to recognize a voice of a specific word uttered by a speaker, variations on a time axis and a frequency axis which are different among speakers are absorbed. To do this, various methods have been proposed.

For example, there are a composite similarity method, a DP matching method on the time axis and the frequency axis, a voice recognition method using a fuzzy theory, and the like.

FIG. 4 shows a block diagram of a system configuration example using the conventional DP matching method. In FIG. 4, 1 is a microphone for receiving a voice and converting it into a voice signal, 2 is a microphone amplifier for amplifying the voice signal, 3 is a band-pass filter group (hereinafter referred to as “filter”) for frequency-dividing the voice signal into a plurality of voice signals. "Bank"), 4 is an A / D converter for converting an analog voice signal into a digital signal, and 5 is voice recognition for performing voice recognition by comparing pre-registered reference registered voice data with voice data to be recognized. Reference numeral 6 is a data registration unit for storing reference voice data.

Next, the operation of the above conventional example will be described. Prior to voice recognition, when voices of an unspecified number of speakers (for example, 20 people) are input from the microphone 1 to a reference specific word, 20 reference voice data are stored in the data registration unit 6. It After that, when a voice to be recognized by an arbitrary speaker is input to the microphone 1, the recognized voice data is sequentially pattern-matched with the 20 reference voice data registered, and the voice recognition unit 5 calculates the similarity. Output is sent.

[0006]

However, the above-described conventional unspecified speaker voice recognition device has a problem that a high recognition rate for a specific word cannot be expected because it is not based on an ideal output. The present invention solves the above-mentioned conventional problems, and provides an excellent speaker-independent voice recognition device that is highly robust against fluctuations in a specific word on the time axis and the frequency axis and that can obtain a high recognition rate. The purpose is to

[0007]

In order to achieve the above object, the present invention analyzes a voice signal corresponding to one or more specific words uttered by an unspecified speaker to generate monitor waveform data. Voice analysis means, one identification system in which monitor waveform data from one of the unspecified speakers is used as identification data and identified by the identification data, and monitor waveform data from a plurality of other speakers are evaluated a plurality of times. Means for generating a plurality of evaluation systems identified by the evaluation data as use data, and a similarity degree estimation value obtained by inputting each evaluation data for the specific word into a corresponding evaluation system,
And a data recognizing unit having a unit for obtaining a least square sum of respective differences from the similarity estimation value obtained by inputting each of the evaluation data to the identification system.

[0008]

Therefore, according to the present invention, the identification data of one specific speaker for a specific word and the evaluation data of a plurality of other speakers are defined, and the similarity estimation based on the identification data and the evaluation data is performed. By recognizing the voice data of a specific word by obtaining the least square sum of the difference between the values, the fluctuations on the time axis and the frequency axis between speakers in each specific word are absorbed,
It is possible to perform speech recognition with high robustness and high recognition rate.

[0009]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a schematic block diagram of an unspecified speaker voice recognition apparatus according to an embodiment of the present invention. In FIG. 1, 11 is a microphone that receives voice and converts it into a voice signal, 12 is a microphone amplifier that amplifies the voice signal, and 13 is a filter bank that frequency-divides the voice signal into multiple (n) voice signals. , 14 are A / D converters for converting a voice signal into voice data of a digital signal. These constitute a voice analysis means for analyzing the voice from the speaker. Reference numeral 15 is a fuzzy identification system that outputs the degree of similarity from the waveform data that has passed through the filter bank 13 (hereinafter referred to as "monitor waveform data"), which constitutes data recognition means.

Next, the creation of the fuzzy identification system will be described. FIG. 1 shows a similarity difference identification system that estimates the similarity of each word to be recognized. Initially, a system for sending the similarity output for each registered word is created in this way. In this case, the number of words to be recognized is 10 numbers from “0 to 9”, and therefore the similarity difference identification system is also 10 systems. Input parameters x1 to xn are supplied to each similarity difference identification system,
The estimated output y'is output.

FIG. 3 shows the configuration of the antecedent and consequent identification systems of each similarity difference identification system. The identification system in this configuration is described in the if-then format, in which the antecedent part is a fuzzy proposition that is a fuzzy variable of a trapezoidal membership function, and the consequent part is an ordinary linear form. is there. If the number of specific words is m (in this case, m = 10), fuzzy rule (hereinafter referred to as “rule”) Ri (i = 1, 2, 3, ..., M)
Becomes m rules, and the fuzzy input xj (j
= 1, 2, 3, ..., N), the fuzzy model shown in (Equation 1) is obtained when the membership function is Ani.

[0013]

[Equation 1] Here, if the fuzzy input xj is a deterministic input (non-fuzzy input) xj0 (x10, x20, ..., xn0), the inference output y'is given by the weighted average by n goodness of fit,
It is represented by (Equation 2) and (Equation 3).

[0014]

[Equation 2]

[0015]

[Equation 3] Yi in (Equation 2) is, as shown in (Equation 4),
This is obtained by substituting the input xj0 into the expression of the consequent part of (Equation 1).

[0016]

[Equation 4] Aji (xj0) is the membership value of the fuzzy variable Aji at xj0, and (Equation 3) is the product of these n pieces. ωi is the product of the goodness of fit of the rule Ri with respect to the inputs x10, x20, ..., Xn0.
It acts as a "weighting" factor when determining.

Since the antecedent part is given by the membership function and the consequent part is given by the linear form in this way, it is described as shown in FIG. 3, for example.

In (Equation 2), [ωi] as shown in (Equation 5) is defined.

[0019]

[Equation 5] By substituting [ωi] and yi shown in (Equation 4) into (Equation 2), (Equation 6) is obtained.

[0020]

[Equation 6] Furthermore, z0i = [ωi], z1i = [ωi] x10, z2i =
Assuming that [ωi] x20, ..., Zni = [ωi] xn0, the output parameter y'is represented by (Equation 7).

[0021]

[Equation 7] Further, since the consequent part is expressed in a linear format, if the antecedent part is determined, the consequent part can be regarded as one linear format. Therefore, the least squares method is used to obtain the estimated value of the unknown constant by the multiple regression analysis method for analyzing the relationship between a plurality of variables. That is, the least squares can be estimated by obtaining the residual between the estimated value and the measured value and minimizing the sum of squares of the residual.
The antecedent part can be obtained by the simplex method of nonlinear programming.

The selection of the input parameters of the fuzzy identification system 15 is performed as follows. It is assumed that the monitor waveform data (x1 to xn) that is considered to be the feature amount of each word (for example, the number of peaks of the waveform data) is selected by referring to a certain other speech recognition and repeating the recognition experiment. In this case, the reference of the voice recognition is, for example, the filter bank 1 of FIG.
It is assumed that the monitor waveform data output from 3 is obtained and the voice recognition result is output according to the similarity.

Table 1 is a data sheet for making the similarity identification system, and the material no corresponds to a code number of a numeral. As the output parameter, as shown in (Table 1), the output value for the target word is as large as 200, and the output value for other words is 0 (or a value close to 0 or a value selected using a random number). And In this case, the same voice analysis system as the device of the present invention is used.

[0024]

[Table 1] That is, this system builds a system with higher word selectivity. After all, this fuzzy identification system
It can be said that the reference speech recognition apparatus is an ideal speech recognition apparatus with improved performance.

As the material data, 10 words of numbers "0-9" generated by a plurality of speakers (for example, 20 people) are used. Further, the identification data is data of one speaker, and the data of the remaining 19 speakers is evaluation data. Usually, in fuzzy identification, the nonuniformity criterion UC used in the GMDH method for modeling the input-output relationship of a nonlinear system with a polynomial of input variables is used. However, in the case of this embodiment, since there are 19 pieces of evaluation data, the UC shown in (Equation 8) is used.

[0026]

[Equation 8] In this (Equation 8), A indicates the system identified by the identification data, and BT indicates the system identified by the evaluation data. For example, yiAJ indicates the similarity estimation value when the data of one number i of the material numbers of the fixed data of A is input to the identification system J.

Next, a plurality of fuzzy identification systems (here, 10 fixed systems) are synthesized. The synthetic identification system in this case is a synthetic identification system having a plurality of similarity degree estimated values in the consequent part with respect to one antecedent part. The composition of the antecedent part is performed by taking a common set of each input parameter.

That is, the fuzzy identification system 15 in the unspecified speaker speech recognition apparatus shown in FIG. 1 estimates 10 similarity degrees of y'1 to y'10 in the consequent part with respect to one antecedent part. It is a synthetic identification system that outputs a value. Therefore, in the recognition operation in FIG. 1, when the speaker generates a specific word (here, the numbers “0 to 9”), the microphone 11, the microphone amplifier 12, the filter bank 13, and A / The monitor waveform data obtained through the D converter 14 is input to the fuzzy identification system 15 which is a data recognition means, and the numeral 10 of the numbers "0 to 9" is input.
The estimated similarity value for each word is output.

[0029]

As described above, according to the present invention, as is apparent from the above-described embodiment, the fuzzy identification system is applied to the unspecified speaker voice recognition device for recognizing a specific word. The effect shown is obtained.

In a specific word, a highly robust recognition operation that absorbs variations on the time axis and the frequency axis between speakers is possible, and it is possible to realize speech recognition with a high recognition rate.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of an unspecified speaker voice recognition device of the present invention.

FIG. 2 is a diagram showing an input / output relationship of a similarity difference identification system for each specific word.

FIG. 3 is a diagram showing a description method of a fuzzy rule of the similarity difference identification system of FIG.

FIG. 4 is a schematic block diagram of an unspecified speaker voice recognition device to which a conventional subtraction method is applied.

[Explanation of symbols]

 11 Microphone 12 Amplifier 13 Filter bank 14 A / D converter 15 Fuzzy identification system

Claims (1)

[Claims] 1. A voice analysis unit that analyzes a voice signal corresponding to one or more specific words uttered by an unspecified speaker to generate monitor waveform data, and a monitor from one of the unspecified speakers. One identification system in which waveform data is used as identification data and identified by the identification data; and a plurality of evaluation systems in which monitor waveform data from other speakers are identified as the plurality of evaluation data by the evaluation data And a similarity estimate value obtained by inputting each of the evaluation data for the specific word into a corresponding evaluation system, and a similarity obtained by inputting each of the evaluation data into the identification system. An unspecified speaker voice recognition device comprising: a data recognition means having a means for obtaining a least sum of squares of respective differences from the degree estimation value.