JP2744622B2 - Plosive consonant identification method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] A plosive consonant discrimination method for discriminating plosive consonants in speech in speech discrimination processing. A closing time detecting means for detecting a closing time point of the input voice at which the vowel preceding the plosive sound ends, a bursting time detecting means for detecting a bursting time of the input voice at which the plosive sound starts, and A closing part near the closing point of the input voice is analyzed with reference to the closing point to obtain a closing part identification parameter, and a rupture part near the rupture point of the input voice is referred to as the puncturing point. Rupture portion analysis means for analyzing as a reference to obtain a rupture portion identification parameter;
Closed part likelihood calculating means for calculating the likelihood between the closed part and each category of plosives from the closed part identification parameter, and determining the likelihood of the plosive part and each category of the plosive sound from the closed part identification parameters It comprises bursting part likelihood calculating means and determining means for determining and identifying the plosive consonant from the likelihood of the closed part and the likelihood of the rupture part.

[Industrial applications]

The present invention relates to a plosive consonant identification method, and more particularly to a plosive consonant identification method for identifying a plosive consonant in a voice in a speech recognition process.

Recognition of plosive consonants is particularly difficult in speech recognition processing, and it is necessary to analyze plosive consonants in input speech in more detail.

[Conventional technology]

FIG. 4 is a block diagram showing a configuration of an example of a conventional plosive consonant identification system.

In the figure, a burst time detecting means 11 detects a burst time of a plosive consonant from a digital time-series signal of an input voice. The analysis means 12 performs an analysis near the detected rupture time point to obtain an identification parameter.

Next, the likelihood calculating means 13 calculates and calculates the likelihood between the identification parameter and each category of the plosive using a standard pattern for each of the preceding and subsequent vowels of the plosive consonant.

Further, the judging means 14 judges the category having the highest likelihood to determine the consonant.

[Problems to be solved by the invention]

In the conventional method, the judgment and discrimination are performed only using the standard parameters obtained by analyzing the vicinity of the burst point, and thus there is a problem that sufficient discrimination performance cannot be obtained depending on the combination of the preceding and succeeding vowels.

For example, in the case of discriminating two of "abi" and "agi", the change of the formant from the preceding vowel "a" to the succeeding vowel "i", that is, the crossing of the formant is affected by the "b"
And the feature of distinguishing between "g" and "g" does not easily appear, and sufficient discrimination performance cannot be obtained.

The present invention has been made in view of the above points, and has as its object to provide a plosive consonant discrimination method having high discrimination performance.

[Means for solving the problem]

 FIG. 1 shows a principle block diagram of the system of the present invention.

In the figure, a digital time-series signal of an input voice is supplied to a closing time detecting means 1 and a rupture time detecting means 2, respectively.

The closing time detecting means 1 detects the closing time of the input voice at which the vowel preceding the plosive consonant ends, and the bursting time detecting means 2 detects the bursting time of the input voice at which the plosive consonant starts.

The closing part analyzing means 3 analyzes a closing part near the closing time point of the input voice with reference to the closing time point to obtain a closing part identification parameter.

The rupture portion analysis means 4 analyzes a rupture portion near the rupture time point of the input voice with reference to the rupture time point to obtain a rupture portion identification parameter.

The closed part likelihood calculating means 5 calculates the likelihood between the closed part and each category of the plosive from the closed part identification parameter.

The plosive part likelihood calculating means 6 calculates the likelihood between the plosive part and each category of the plosive from the plosive part identification parameter.

The determining means 7 determines and identifies a plosive consonant from the likelihood of the closed part and the likelihood of the rupture part.

[Action]

In the present invention, in addition to the plosive part where the plosive consonant starts, the characteristics of the closed part where the preceding vowel ends are extracted, the likelihood of each category is obtained for each, and the highest likelihood is obtained by integrating them. Determine the category.

As a result, for example, each of "abi" and "agi", which were difficult to identify only by the rupture portion, can be identified, and the identification performance is improved.

〔Example〕

FIG. 2 is a block diagram showing the configuration of an embodiment of the system of the present invention. In this embodiment, a word voice including a voiced consonant “b, d, g” is input. The discrimination between the voiced plosive consonants is performed.

In the figure, reference numeral 20 denotes a voice data memory, which is a voiced plosive, a preceding vowel and a succeeding vowel in FIG.
The digital time series signal of the input voice as shown in FIG. Note that the input voice in FIG.
"i" is pronounced.

The digital time-series signal read from the audio data memory is supplied to the closing time detecting unit 21 and the rupture time detecting unit 22, respectively.

The closing point detecting section 21 detects, as the closing point A, a point in time when the power of the input sound with the high-frequency emphasis becomes lower than a certain threshold value TH1 as shown in FIG. 3 (B). The closing point A is a point at which the articulation points of the lips, tongue, palate, etc. are closed and the preceding vowel section ends.

Further, the rupture time detection unit 22 detects a time point when the power of the input voice shown in FIG. 3B becomes higher than a certain threshold value TH2 as a rupture time point B. The burst point B is a point at which the articulation point is released and the consonant section starts.

If then closure analysis position setting unit 23 for performing an analysis of three frames, the analysis frame T ₀ and the analysis frame T _-1 preceding it to the closure point A the analysis frame period as T as a reference (center), T _- Set ₂ .

Thereafter, the closing part analyzing unit 24 performs a frequency analysis of the analysis frame T0, for example, divides a frequency of ₀ to 8 kHz into 16 bands, obtains a power spectrum of each of the 16 bands, and similarly analyzes the analysis frame T _−1. , T _-2 are also _subjected to frequency analysis to obtain a total of 48 power spectra, which are used as 48-dimensional closed part identification parameters X (elements x _i , i = 1 to 48).

By the way, the closed part standard pattern dictionary 26 has a principal component coefficient vector M
And "b" of the eight-dimensional data expanded by the principal component,
The average vectors Eb, Ed, Eg of the groups “d” and “g”, respectively,
The inverse matrix V of the average variance-covariance matrix of the variance-covariance matrix of each group of “b”, “d”, and “g” is stored.

Here, the principal component expansion is to compress a 48-dimensional parameter into an eight-dimensional parameter, and a principal component coefficient vector M is a vector of coefficients necessary for this principal component expansion, and is 8 × 8
It is a dimensional matrix (elements m _ij , i = 1 to 48, j = 1 to 8).
The average vector E (Eb, Ed, Eg) of each group is an 8-dimensional vector (elements e _j , j = 1 to 8), and the variance-covariance matrix represents the variance of a large number of data in each group. A matrix, V
Is an 8 × 8-dimensional matrix (elements v _ij , i = 1 to 8, j = 1 to 8).

The closed part principal component developing unit 25 calculates the following equation from the closed part identification parameter X obtained by the closed part analysis unit 24 and the principal component coefficient vector M from the closed part standard pattern dictionary 26, and calculates the main Find component R (element r _j , j = 1 to 8) Next, the closing portion distance calculation unit 27 performs the following calculation from the above-described principal component R, the average vectors Eb, Ed, Eg and the inverse matrix V from the closing portion standard pattern dictionary 26, and calculates each category “b”,
The distance P (scalar amount) between “d” and “g” is obtained.

P _q = (R−E _q ) ^t · V · (R−E _q ) (2) (where q = b, d, g, and t represent transpositions) Each of the distances Pb, Pd, and Pg is The smaller the likelihood with each category, the smaller the value.

Next, when analyzing three frames, the rupture portion analysis position setting unit 28 sets the analysis frame period T as the analysis frame T ₁ with the rupture time point B as a reference (center) and the analysis frames T ₂ and T ₃ preceding the analysis frame T ₁ . Set.

Thereafter, the rupture part analyzing unit 29 is divided in the band of the analysis frame T ₁ of the performs frequency analysis for example the frequency 0～8KHz 16, obtains the power spectrum of s the 16 bands each similarly analyzed frame T _2, by performing frequency analysis also T _3, obtains a total of 48 power spectrum, which is referred to as 48-dimensional rupture portion identification parameter Y (component y _{i, i} = 1~48).

By the way, in the rupture part standard pattern dictionary 31, the principal component coefficient vector N of the rupture part obtained in advance from a large number of data is stored.
And "b" of the eight-dimensional data expanded by the principal component,
Average vectors Fb, Fd, Fg of the respective groups of “d” and “g”;
The inverse matrix W of the average variance-covariance matrix of the variance-covariance matrix of each group of “b”, “d”, and “g” is stored.

Here, the principal component coefficient vector N is a vector of coefficients necessary for principal component expansion, and is an 8 × 8 dimensional matrix (element n _ij , i = 1
4848, j = 1１〜8). Also, the average vector F of each group
(Fb, Fd, Fg) is an 8-dimensional vector (elements f _j , j = 1 to 8)
And W is an 8 × 8 dimensional matrix (elements w _ij , i = 1 to 8, j =
1 to 8).

The rupture portion main component developing unit 30 includes a rupture portion identification parameter Y,
Principal component coefficient vector N from rupture part standard pattern dictionary 31
And the following equation is used to obtain the principal component S (element s _j , j = 1 to 8) of the parameter Y Next, the rupture portion distance calculation unit 32 performs the following calculation from the above-described principal component R, the average vectors Fb, Fd, Fg and the inverse matrix W from the rupture portion standard pattern dictionary 31, and obtains each category “b”,
The distance Q (scalar amount) between “d” and “g” is obtained.

Q _q = (S−F _q ) ^t · W · (S−F _q ) (4) (however, q = b, d, g) Each of the distances Qb, Qd, Qg is the likelihood of each category. The smaller the value, the smaller the value.

The determining unit 33 includes a closing unit distance calculating unit 27 and a rupture unit distance calculating unit 32.
The closing part distance P and the rupture part distance Q of each category supplied from each are added, and the category having the minimum value and the highest likelihood is determined, and this is determined as the identification result.

As described above, in addition to the rupture part, the characteristics of the closed part are captured, and comprehensive judgment is performed from the likelihood of each part. can get.

〔The invention's effect〕

As described above, according to the plosive consonant discrimination method of the present invention, plosive consonants, which were conventionally difficult to discriminate depending on the combination of the preceding and succeeding vowels, can be distinguished and high discrimination performance is obtained, which is extremely useful in practice.

[Brief description of the drawings]

1 is a block diagram showing the principle of the system of the present invention, FIG. 2 is a block diagram showing the configuration of an embodiment of the system of the present invention, FIG. 3 is a diagram showing an input voice waveform and its logarithmic power, and FIG. FIG. 3 is a block diagram illustrating a configuration example of a system. In the figure, 1 is a closing time detecting means, 2 is a rupture time detecting means, 3 is a closing part analyzing means, 4 is a rupture part analyzing means, 5 is a closing part likelihood calculating means, 6 is a rupture part likelihood calculating means, 7 Indicates a judgment means.

Claims (1)

(57) [Claims] 1. A plosive consonant discrimination method for identifying a plosive consonant in an input voice, comprising: a closing time detecting means (1) for detecting a closing time of the input voice at which a vowel preceding the plosive consonant ends; A burst point detecting means (2) for detecting a burst point of the input voice at which a consonant starts, and a closing part near the closing point of the input voice is analyzed with reference to the closing point to obtain a closing part identification parameter. Closing part analyzing means (3); rupture part analyzing means (4) for analyzing a rupture part near the rupture time point of the input voice with reference to the rupture time point to obtain a rupture part identification parameter; Closing part likelihood calculating means (5) for obtaining the likelihood between the closing part and each category of plosives from the parameters; and bursting for obtaining the likelihood between the plosive part and each category of the plosives from the burst identification parameters. Likelihood And arithmetic means (6), rupture consonant detection method, characterized in that from the likelihood of the likelihood and the rupture of the closure having a determining identifying judging means (7) said burst consonants.