JPH03239299A - Fricative consonant discriminating system - Google Patents

Abstract

PURPOSE:To obtain a high discrimination performance by deriving a discrimination parameter from both a consonant section and the succeeding vowel transition part. CONSTITUTION:A consonant section analyzing means 2 analyzes the inside of a consonant section and derives a discrimination parameter in the consonant section. A succeeding vowel transition part analyzing means 6 analyzes the succeeding vowel transition part extending from a consonant section end point to the succeeding vowel rise point, and derives a discrimination parameter in the succeeding vowel transition part. A likelihood arithmetic means 3 derives a likelihood of an input voice and each category of a fricative consonant from the discrimination parameter in the consonant section and the discrimination parameter in the succeeding vowel transition part, and a deciding means 4 decides the category whose derived tolerance is the largest and decides and discriminates a consonant. In such a way, the fricative consonant whose discrimination had been difficult by only a consonant part can be discriminated, and the discrimination performance can be improved.

Description

[Detailed Description of the Invention] [Table of Contents] Overview Industrial Application Fields Prior Art and Problems to be Solved by the Invention Means for Solving the Problems Action Examples Effects of the Invention [Summary] In voice recognition processing, voice Regarding the fricative consonant identification method that identifies middle fricative consonants, with the aim of achieving high discrimination performance, we calculated the identification parameter ■ not only from the consonant interval but also from the transition part of the preceding and following vowels, and further, we and analysis conditions, e.g. analysis window period (F).

It is configured to change the analysis window length (W).

(Industrial Application Field) The present invention relates to a method for identifying fricative consonants, which is important in speech recognition processing.

With recent advances in voice recognition technology, efforts are being made to put voice word processors and voice-based data input in factories into practical use.

In these fields, data is input by voice, so a high recognition rate for the human-generated voice is required.

Generally, recognition of vowels and consonants is important for accurate speech recognition, and in speech recognition, it is often easy to recognize vowels, but difficult to recognize consonants.

Therefore, it is important to improve the consonant recognition rate for the practical application of voice data input.

[Prior art and problems to be solved by the invention] FIG. 6 is a diagram showing the configuration of a conventional fricative consonant identification system.

In the figure, a consonant section detecting means 1 detects the starting point of a consonant section from a digital time-series signal of human speech.

And detect the end point.

The analysis means 2 performs analysis within the detected consonant section to obtain identification parameters.

Next, the likelihood calculation means 3 uses the standard pattern of fricative consonants to calculate the likelihood of the discrimination parameter ■ and each category of fricative consonants (i.e., consonants "φ", "S", and "h"). I ask.

Further, the determining means 4 determines the category with the highest calculated likelihood and performs determination and identification of fricative consonants.

In the above conventional method, judgment and identification are performed only using the identification parameters obtained by analyzing the inside of the consonant interval, so even if there is information effective for identification in the transition part of the following vowel and preceding vowel, it cannot be used. However, there was a problem in that sufficient identification performance could not be obtained.

SUMMARY OF THE INVENTION In view of the above-mentioned conventional drawbacks, it is an object of the present invention to provide a fricative consonant identification method that has high discrimination performance for identifying fricative consonants in input speech.

[Means to solve the problem]

FIG. 1 shows a block diagram of the principle of the present invention, in which (a) shows the case where the identification parameter is determined from both the consonant interval and the subsequent vowel transition part, and (b) shows the case where the identification parameter is determined from both the preceding vowel transition part and the consonant interval. A case is shown in which the identification parameter is determined from each of the following vowel transition parts.

The above problems are solved by a fricative consonant identification method configured as follows.

(1) In the fricative consonant identification method for identifying fricative consonants in input speech, the identification parameter is determined from both the consonant interval and the subsequent vowel transition part.

(2) In the above fricative consonant identification method, different analysis window periods (F) are used in the consonant interval and the subsequent vowel transition part.

(3) In the above fricative consonant identification method, different analysis window lengths (W) are used for the consonant interval and the subsequent vowel transition part.

(4) In the fricative consonant identification method for identifying fricative consonants in input speech, the identification parameter is determined from three locations: the preceding vowel transition part, the consonant interval, and the following vowel transition part.

(5) In the fricative consonant identification method described in (4) above,
The preceding vowel transition section, the consonant section, and the following vowel transition section are configured to use different analysis window periods (F), respectively.

, 5 (6) In the fricative consonant identification method described in (4) above,
Different analysis window lengths (-) are configured to be used for the preceding vowel transition part, the consonant section, and the following vowel transition part.

(Operation) That is, according to the present invention, in order to identify a fricative consonant, which is important in speech recognition processing, instead of determining the identification parameter from the consonant interval as in the past, the identification parameter is determined from the consonant interval and before and after the consonant. The identification parameter is also determined from the transition part of the vowel.

First, a digital time-series signal of input speech is supplied to the consonant interval detection means 1 and the subsequent vowel rising point detection means 5.

The consonant interval detection means 1 detects the start point and the end point of the consonant interval, and the subsequent vowel rising point detection means 5 detects the point at which the transition from the consonant to the subsequent vowel ends and the utterance of the subsequent vowel begins to be determined. Detect.

The consonant section analysis means 2 analyzes the inside of the consonant section to obtain identification parameters in the consonant section.

The subsequent vowel transition analysis means 6 analyzes the subsequent vowel transition from the end point of the consonant section to the rising point of the subsequent vowel, and obtains the identification parameter (2) for the subsequent vowel transition.

The likelihood calculation means 3 calculates each category of human speech and fricative consonants (i.e., ``φJ 'Sj 'h
J) Find the likelihood.

The determining means 4 determines the category with the greatest likelihood and performs consonant identification.

Furthermore, if there is a wealth of data used to create a dictionary, a dictionary can be created for the preceding vowel and subsequent vowel, and information can also be obtained from the transition part of the preceding vowel. The principle diagram in this case is shown in Figure 1 (
Shown in b).

The preceding vowel transition start point detecting means 7 detects the point at which the transition from the preceding vowel to the consonant begins.

The preceding vowel transition part digit means 8 analyzes the preceding vowel transition part from the starting point of the preceding vowel transition to the starting point of the consonant section, and obtains the identification parameter (2) for the preceding vowel transition part.

The likelihood calculation means 3 calculates the likelihood between the input speech and each category of fricative consonants from the identification parameters at the preceding vowel transition part, the identification parameters at the consonant interval, and the identification parameters at the subsequent vowel transition part. .

Judgment means 4 selects the category (
Consonant) is determined and the consonant is determined and identified.

In this way, in the present invention, features are extracted not only from the consonant interval but also from the transition part between the preceding vowel and the following vowel.
The likelihood with each category is determined using all the identification parameters, and the category with the highest likelihood is taken as the identification result.

This makes it possible to identify fricative consonants, which were difficult to identify using just the consonant part, and has the effect of improving identification performance.

〔Example〕

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

The above-mentioned FIG. 1 is a block diagram of the principle of the present invention, FIG. 2 is a diagram showing one embodiment of the present invention, and FIG. 3 is a diagram showing another embodiment of the present invention. FIG. 4 is a diagram showing an example of analysis conditions for a consonant interval and a vowel transition part according to the present invention, where (a) shows an example of a consonant interval, and (b) shows an example of a subsequent vowel transition part. (c) shows an example of the preceding vowel transition part, and FIG. , find the identification parameters not only from the consonant interval but also from the transition part to the vowel before and after the consonant, and analyze the consonant interval and the analysis conditions at the vowel transition part, for example, the period of the analysis window (F
) 1 analysis window length (The means for changing the old is a necessary means for carrying out the present invention. The same reference numerals indicate the same objects throughout the figures.

Hereinafter, while referring to FIG. 1, the fricative consonant identification method of the present invention will be explained with reference to FIGS. 2 and 3.

Figure 2 shows claim (1) of the fricative consonant identification method according to the present invention.
This shows the configuration of ~(3).

In this embodiment, the speech of a word containing the voiceless fricative consonants ``φ, s, h, and 1'' is manually generated, and the voiceless fricative consonants are discriminated.

In the figure, the audio data memory is supplied with and records digital time-series signals of input audio of voiced/voiceless fricative 0 consonants, vowels preceding them, and vowels following them.

The digital time-series signal read from the audio data memory is sent to the consonant section start point detection section 10.

The signal is supplied to the consonant section end point detecting section 11 and the subsequent vowel rising point detecting section 5, respectively.

The consonant section start point detection unit 10 detects the point in time when the low frequency power of the input voice becomes lower than a certain threshold value THI as the consonant section start point. The consonant section end point detection unit 11 detects the point in time when the low frequency power of the human voice becomes higher than a certain threshold TH2 as the consonant section end point.

Further, the subsequent vowel rising point detection unit 5 detects the point in time when the low frequency power of the input voice becomes higher than a certain threshold TH3 (>TH2) as the subsequent vowel rising point. (Figure 4,
(See FIG. 5) Next, when analyzing three frames, the consonant section analysis position setting unit 20 divides the space between the consonant section start point and the consonant section end point into four equal parts (thereby, the analysis period (F ) is determined (the same applies hereinafter), and an analysis frame with a window length W1 is set at the three central dividing points.

(See FIG. 4(a)) After this, the consonant interval analysis unit 21 performs frequency analysis of the analysis frame, for example, divides the frequency from 0 to 10 kHz into 32 bands, and calculates the power spectrum of each of the 32 bands. , and a total of 96 power spectra are identified by the identification parameter X (element χ.

i=1 to 96).

Next, the subsequent vowel transition part break position setting unit 60, for example,
When analyzing three frames, the area between the consonant interval end point and the following vowel start point is divided into two, and an analysis frame with a window length W2 is set at the three division points. (See FIG. 4(b)) After this, the subsequent vowel transition part splitting unit 61 performs frequency analysis of the analysis frame, for example, frequency O~10kHz.
is divided into 32 bands, the power spectra of each of these 32 bands are determined, and a total of 96 power spectra are determined by the identification parameter X (element Xi).

1-97~192) ■.

By the way, the standard pattern dictionary 31 contains the principal component coefficient vector M of the identification parameter ■, which has been obtained in advance from a large amount of data, and "φ" of the 28-dimensional data expanded with two principal components.
, "S", "h", each group's average vector Eφ, Es,
Eh and the inverse matrix (■) of the mean variance-covariance matrix of the variance-covariance matrix of each group are stored.

Here, principal component expansion means compressing the above 192-dimensional parameters into 28-dimensional parameters without losing the amount of information, and the principal component coefficient vector M is the coefficients necessary for this principal component expansion. is a 192x28-dimensional matrix (element m).

i=1-192. j=1 to 28). In addition, the average vector E (Eφ, Es, Eh) of each group is a 28-dimensional vector (elements e, , j = 1 to 28), and the variance-covariance matrix is a vector of many data (elements 〉, where ■ is a 28×28-dimensional matrix (element V i j
+ 1-1~28. j-1 to 28).

Next, the principal component analysis unit 30 calculates the following equation from the identification parameter Then, the principal component R (element rJ, j=1 to 28) of the parameter X■, that is, the compressed data is obtained.

rj−Σm,,−X, (j = 1 to 28)
-------m---(1) Next, the distance calculation unit 32 calculates the above principal component R, the average vectors Eφ, Es, Eh from the standard pattern dictionary 31, and the inverse matrix V (the above, each When calculating the distance P to each consonant, which will be described later, based on the variance of the elements, division is usually required to normalize each variance, but the inverse matrix of the variance is used. (By doing so, the division can be reduced to multiplication), calculate the distance P from each category "φ" 5 "s" 1 "h" by calculating the following formula.
Find (scalar quantity).

pq−(R−Eq)t −v・(R−Eq)(2) (However, in 9−φ, s, h, L represents transposition) This distance Pφ, Ps, Ph is the distance between each category. 3 4 The higher the likelihood, the smaller the value.

Regarding the derivation process of formula (1) for calculating the principal component R, see, for example, "Multivariate Analysis Method, P, 84-100゛Principal Component Analysis Method," by Haruo Yanai and Yoshio Takashi, Asakura Shoten, 1.
977, first published on September 30th, so I will only show the calculation formula. Also, find the above distance P (2)
Regarding the process of deriving
, ``Discriminant Analysis'', P. 69'', so only the calculation formula will be shown.

The determining unit 4 determines the category with the lowest value of distance P for each category supplied from the distance calculating unit 32 and the highest likelihood, and uses this as the identification result.

Furthermore, if the speech data for creating the standard pattern dictionary 31 is very rich, it is possible to create dictionaries for preceding and succeeding vowels and obtain identification parameters from the transition part of the preceding vowel.

FIG. 3 is a block diagram showing the configuration in that case, and it shows claims (4) to 3 of the fricative consonant identification method according to the present invention.
This corresponds to (6). This example also uses the voiceless HM consonant “φ,
A word sound including "s, t+" is input, and the voiceless fricative consonants are discriminated.

In this case, the digital time-series signal read from the audio data memory is sent to the preceding vowel transition start point detection section 10, the consonant section start point detection section 10, the consonant section end point detection section 1], and the subsequent vowel transition detection section 1. Outputs 54)-4 are supplied, respectively.

The preceding vowel transition start point detection unit 7 detects the point in time when the low frequency power of the input voice becomes lower than a certain threshold THO as the preceding vowel transition starting point. The consonant section start point detection unit 10 detects a dark value THI in which the low frequency power of the input voice from which the high frequency has been removed is constant.
(<THO) The time point when the value becomes lower than THO is detected as the consonant section start point. The consonant section end point detection unit 11 detects the point in time when the low frequency power of the input voice becomes higher than a certain threshold value TH2 as the consonant section end point. Further, the subsequent vowel rise point detection unit 5 detects a threshold value TH at which the low frequency power of the input voice is constant.
3 (>TH2) is detected as the trailing vowel rising point. (See FIGS. 4 and 5) Next, the preceding vowel transition part break position setting unit 80, for example,
When analyzing three frames, the space between the preceding vowel transition start point and the consonant section start point is divided into two equal parts, and analysis frame 1 with window length W0 is set at the central three division points. (Figure 4(c)
) After this, the preceding vowel transition part fragmentation unit 8I performs a frequency analysis of the above analysis frame, for example, the frequency of 0 to 10kHz.
is divided into 32 bands, the power spectrum of each of these 32 bands is obtained, and a total of 96 power spectra are used as the identification parameter Y (elements y8.i-1 to 96).

Next, when analyzing three frames, for example, the consonant section analysis position setting unit 20 divides the space between the consonant section start point and the consonant section end point into four equal parts, and sets the window length to the three central dividing points. W1
Set up an analysis frame.

(See FIG. 4(a)) After this, the consonant interval analysis unit 21 performs a frequency analysis of the analysis frame, for example, divides the frequency from 0 to 10 kHz into 32 bands, and analyzes each of the 32 bands. A total of 96 power spectra are determined using the identification parameter Y (element y.

1-97 to 192).

Next, the subsequent vowel transition part break position setting unit 60, for example,
When analyzing three frames, divide the area between the consonant interval end point and the following vowel rise point into two, and set the window length W at the three dividing points.
Set up the second analysis frame. (See FIG. 4(b)) After this, the subsequent vowel transition part splitting unit 61 performs frequency analysis of the above analysis frame, for example, the frequency of 0 to 10kHz.
is divided into 32 bands, the power spectrum of each of these 32 bands is determined, and a total of 96 power spectra are used as the identification parameter Y (element yt,!=193 to 28 days).

By the way, the standard pattern dictionary 31 contains the principal component coefficient vector N of the identification parameter ■ previously obtained from a large amount of data, and the principal component coefficient vector N of the 28-dimensional data expanded as the principal component.
”, “S”, “h” group mean vectors Fφ, Fs
, Fh, and the inverse matrix U of the mean variance-covariance matrix of the variance-covariance j-te sequence of each group are stored.

7 8 Here, principal component expansion means compressing the above 288-dimensional parameters into 28-dimensional parameters without losing the amount of information, and the principal component coefficient vector N is the number of coefficients required for this principal component expansion. 288x28-dimensional vector matrix (element n ill ”” 1 to 28B, j = 1 to 28)
It is. Also, the average vector F(Fφ, FsSFh
) is a 28-dimensional vector (element fJ, j = 1 to 28), the variance-covariance matrix is a matrix representing the variance of a large number of data in each group, and U is a 28x28-dimensional matrix (element uij , i""l~28.j=1~28).

The principal component analysis section 30 uses the identification parameter Y obtained by the preceding vowel transition section folding section 81, the consonant interval analyzing section 21, and the following vowel transition section folding section 61, and the standard pattern dictionary 31.
The following equation is calculated from the principal component coefficient vector N from
seek.

The following equation is calculated from the average vectors Fφ, FsFh from the pattern dictionary 31 and the inverse matrix U, and each category 1φ1.
Find the distance Q (scalar quantity) between 'SJ and 'hJ.

Qq−(S−Fq)L −U・(S−Fq)(2) (However, U5, +4 = φ, s, h, and ′ represents transposition) This distance study φ Qs, Qh is The likelihood of is as small as the quotient.

The determination unit 4 determines the category with the lowest value of the distance Q of each category supplied from the 1-record distance calculation unit 32 and the highest likelihood, and uses this as the identification result.

In this way, the present invention analyzes not only the consonant interval but also the transition part between the preceding and succeeding vowels under optimal analysis conditions (period (F), window length (W)) and determines the likelihood. The feature is that it is possible to obtain higher identification performance than the conventional method.

〔Effect of the invention〕

As described above, according to the fricative consonant identification method of the present invention, in speech identification processing, identification parameters are determined not only from the consonant interval but also from the transition parts of the preceding and following vowels, and furthermore, the identification parameters are determined from the consonant interval and the preceding and following vowel transitions. The analysis conditions (analysis window period (F
) 1 analysis window length (h)), high discrimination performance can be obtained and the effect is that it is extremely useful in practice.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the principle of the present invention. While FIG. 2 shows one embodiment of the invention, FIG. 3 shows another embodiment of the invention. Figure. FIG. 4 is a diagram showing an example of analysis conditions for a consonant interval and a transition part of a vowel according to the present invention. Figure 5 is a diagram showing an example of detecting the consonant 1g interval and the vowel transition part using the low-frequency power and darkness value of the human voice. Figure 1: 1! 〆] Figure C shows the configuration of a bundle fricative consonant identification system. It is. In the drawing, 1 is a consonant interval detection means. 10 is a consonant section start point detection section 11 is a consonant section end point detection section. 2 is a first stage of analysis or a consonant interval analysis means. 20 is a consonant interval analysis position setting unit. 21 is a child j) The interval analysis section 3 is a likelihood calculation means. 30 is the power analysis department. Reference numeral 31 denotes a standard pattern dictionary; 32, a distance calculation section 4 is a determining means; or, a determining section 5 is a subsequent vowel rising point detecting means 1; or a subsequent vowel rising point detecting section 6 is a subsequent vowel transition partial digit means. 60 is a subsequent vowel transition part break position setting unit. Reference numeral 61 denotes a subsequent vowel a transition part analysis unit 7, which is a preceding vowel transition start point detection means or a preceding vowel transition start point detection unit. 1 2 8, the preceding vowel transition part analysis means 80 is a preceding vowel transition part breaking position setting unit. 81 is the leading vowel transition part. ■ is an identification parameter. M, N are the principal component coefficient vectors Eφ, Es, Eh, Fφ, Fs, Fh, are the mean vector V of the 7-tones, and tJ is the inverse matrix of the mean variance-covariance matrix of the variance-covariance matrix. S and R are principal components, P and Q are distances. are shown respectively. 3 Principle block diagram of the present invention Figure 1 (Part 1)

Claims (6)

[Claims] (1) In the fricative consonant identification method that identifies fricative consonants in input speech, the identification parameter (
(1)) A fricative consonant identification method characterized by determining the following. (2) In the above fricative consonant identification method, different analysis window periods (F
2. The fricative consonant identification method according to claim 1, wherein the method uses: ). (3) In the above fricative consonant identification method, the analysis window length (W) is different between the consonant interval and the following vowel transition part.
2. The fricative consonant identification method according to claim 1, wherein: (4) In a fricative consonant identification method for identifying fricative consonants in input speech, a fricative consonant characterized in that the identification parameter ((1)) is obtained from three locations: a preceding vowel transition part, a consonant interval, and a subsequent vowel transition part. Identification method. (5) In the fricative consonant identification method according to claim 4, the fricative consonant identification method is characterized in that different analysis window periods (F) are used for the preceding vowel transition part, the consonant interval, and the following vowel transition part. . (6) The fricative consonant identification method according to claim 4, wherein different analysis window lengths (W) are used for the preceding vowel transition part, the consonant interval, and the following vowel transition part.