JPS6325699A - Formant extractor - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a formant extraction device used for speech analysis and recognition.

BACKGROUND OF THE INVENTION A formant is a peak of resonance that occurs on the frequency spectrum of a speech wave due to the poles of the transfer function of the vocal tract, and is an important factor that shapes phonology. 1st in order from the lowest frequency. Second... called formant.

Conventionally, various methods have been used for formant extraction, such as reading Tunagrams, using a filter bank, and analyzing by synthesis.

This method also includes a method using linear predictive analysis:
This is a close approximation of the vocal tract transfer function using an all-pole model.

An example of the conventional formant extraction device mentioned above will be described below with reference to the drawings. FIG. 2 is a block diagram of the main parts of a conventional formant extractor. In the same figure, 2
1 is a linear prediction analysis section, 22 is a high-order equation root finding section, and 23 is a formant selection section.

The operation of the conventional formant extraction device configured as described above will be described below.

The third diagram is a processing flow diagram in the linear prediction analysis unit 21. Here, the audio sampling frequency is 10Kllz
, ■The time length of the frame is 20 m5. One frame of audio waveform data is expressed as y (i), i=1 to 200.

For pre-emphasis, a first-order difference expressed by the following equation is used.

y'(i>=y (i+1)-y (i)
--- (1) In order to reduce the effect of frequency distortion caused by frame extraction, a Hamming window is applied as shown by the following equation.

y′ (i)=H(i) ・y′ (i) −・−
・−(2)H(i) −〇, 54−○,, +6 CO5(2π i /200) −1−−<
3+Next, calculate the short-term autocorrelation function R1 expressed by the following equation.

1111・-(4) The α parameter is obtained by solving the following simultaneous -order equations.

In reality, it is possible to efficiently solve the problem using a recursive solution method such as the Durbin method, and the k parameter can also be determined at the same time.

The transfer function of the vocal tract is expressed by the following equation.

Therefore, in the higher-order equation root finding section 22, 1+Σ α
, Z-"=0 by solving the roots of the higher-order equation using the Newton'-Labran method, etc., the poles of the vocal tract transfer function can be found.Furthermore, in the formant selection section 23, the frequency and bandwidth are taken into consideration. Then select the root corresponding to the desired formant from among multiple roots.

Problems to be Solved by the Invention In the formant extracting device configured as described above, there is a problem in that floating point calculations are required and processing time is required because high-order algebraic equations must be solved. Also, errors occur when selecting a root corresponding to the desired formant from among the multiple roots found.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a formant extraction device that can easily and uniquely obtain a desired formant frequency.

Means for Solving the Problems In order to solve the above-mentioned problems, the present invention includes a vowel discriminator that discriminates vowels in input vowel speech waveform data based on speech feature quantities, and a vowel discriminator that discriminates vowels in input vowel speech waveform data using speech feature quantities, and the above-mentioned speech characteristics of corresponding plural vowels of plural speakers in advance. Based on the volume data and formant frequency data, multiple regression analysis is performed to estimate the formant frequency from the voice feature amount for each gender and each vowel, and the regression coefficients obtained are stored.
A formant estimation coefficient storage unit that outputs a regression coefficient for the corresponding vowel and gender based on the vowel discrimination result and the gender information in the human voice waveform data, and a formant estimation coefficient storage unit that performs regression from the voice feature amount and the regression coefficient for the corresponding vowel and gender. and a formant estimator that calculates and outputs a formant frequency estimated value as a linear value.

Effects of the present invention As described above (1'N formation, vowel discrimination and gender 48 information, the corresponding one is selected from among the regression coefficients prepared for each gender and each vowel, and the voice feature amount and this regression The formant frequency can be determined by a product-sum operation with the coefficients.

The multiple monthly analysis method consists of a certain variable y and other variables that are thought to affect it, X+, Xz+...X.

Prediction based on data regarding ao +3. X, “”+aF XF y
This is one way to predict. The accuracy of prediction depends on the degree of correlation between variables x, l xz + . . . , and variable y (degree of fit to the regression line). Also, if the data fit poorly to the regression line calculated from N pieces of data, the variables X + + Xz + ...X
, by dividing the space of , into smaller data sets (number of data n
1. By performing prediction every n2...<<N), accuracy can be improved. In the present invention, by gender,
Regression coefficients are calculated for each vowel.

Here, the method for calculating the regression coefficient will be explained.

P-order audio features ixi. , i=l, . . . , from P to the j-th formant frequency f, , j=1. The regression model for estimating . . . is expressed by the following equation.

At this time, the estimation error e. is expressed by the following equation.

Let ε2 be the sum of N samples e, , ′ covering multiple vowels from multiple speakers.

Condition to minimize ε2 δε2/δa4, -〇(Q<i<
P), by solving the following coalition - next step 2,
Regression coefficients can be found.

EXAMPLE Hereinafter, a formant extracting apparatus according to an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of essential parts showing a first embodiment of the formant extracting apparatus of the present invention. In FIG. 1, 11 is a sound noise analysis section, 12 is a vowel discrimination section, 13 is a formant estimation coefficient storage section, and 14 is a formant estimation section.

Regarding the formant extraction device configured as above,
The operation will be explained below.

First, the sound volume analysis unit 11 analyzes input speech waveform data to obtain a P-dimensional speech feature lxi representing a spectral envelope.
, i = 1. ...Calculate P.

Based on the data samples obtained for each vowel regarding vowel discrimination and (the amount of phonetic respect xi, i = 1...P),
This is a method of determining which vowel the input speech waveform data belongs to based on these speech values, and a linear discriminant function is often used.

1st. Assuming that two formant frequencies of the second formant are to be estimated, the formant estimation coefficient storage unit 13 stores in advance the voice feature data of the five Japanese vowels of multiple male and female speakers and the first and second formant frequencies. Based on the second formant frequency data, the first formant frequency data is calculated from the voice features for each gender and each vowel. Write down the regression coefficients obtained by performing multiple regression analysis to estimate the second formant frequency. Then, based on the vowel sentence result and gender information, regression coefficients for the corresponding vowel and gender are output. This is aji, J=1゜2i
= o, . . . shall be represented by P.

In the formant estimating unit 14, the voice feature lx1
．． i=1. −P and the above regression coefficient “i;, J−1,2i=
0. . . . From P, calculate the 1st degree second formant period/i number quasi-definite value f;, J=1.2 using the following equation.

As described above, according to the tree embodiment, by providing the vowel discrimination unit 12, the formant estimation coefficient storage unit 13, and the formant estimation unit 14, the speech time ff representing the spectral envelope is
From 1ffi, the formant frequency can be estimated.

A second embodiment of the present invention will be described below with reference to the drawings.

FIG. 4 is a block diagram of the main parts of a formant extraction device showing a second embodiment of the present invention. The configurations of the vowel discrimination section 42, formant estimation coefficient storage section 43, and formant estimation section 44 are similar to those shown in FIG. A linear predictive analysis section 41 is used as the abutment analysis section, and the α parameter is output as a phonetic trace representing the spectrum envelope.

The processing in the linear prediction analysis unit 41 is as shown in the processing flow diagram of the linear prediction analysis unit 21 in the conventional example shown in FIG.
(Figure).

By providing the vowel discrimination section 42, formant estimation coefficient storage section 43, and formant estimation section 44, the amount of calculation can be significantly reduced compared to the conventional method of solving a high-order algebraic equation to obtain the formant frequency. Can be done.

A third embodiment of the present invention will be described below with reference to the drawings.

FIG. 5 is a block diagram of main parts of a formant extracting apparatus showing a third embodiment of the present invention. The configurations of the vowel discrimination section 52, formant estimation coefficient storage section 53, and formant estimation section 54 are the same as those shown in FIG. A linear predictive analysis unit 5 that calculates the α parameter from the audio waveform data input to the acoustic analysis unit.
11, and a cepstrum calculation unit 512 that calculates LPC cepstrum coefficients from the α parameter.

The processing in the linear prediction analysis section 511 is the same as the processing flow diagram (FIG. 3) in the linear prediction analysis section 21 in the conventional example shown in FIG.

In the cepstrum calculation unit 512, L
PC cepstral coefficient C;, i=1. ...P--1--Cl21 Effects of the Invention As described above, the present invention provides a vowel discrimination section, a formant estimation coefficient storage section, and a formant estimation section, thereby determining the phonetic trace amount and formant frequency representing the spectral envelope. When linear predictive analysis is used in the acoustic analysis section and the α parameter is used as the phonetic trace, it is possible to estimate the , the amount of calculation can be significantly reduced. In addition, a linear prediction analysis section and a cepstrum calculation section are used in the acoustic analysis section, and the LPC is used as the phonetic trace amount.
A similar effect can be obtained by using cepstral coefficients.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of main parts of a formant extractor according to the first embodiment of the present invention, Fig. 2 is a block diagram of main parts of a conventional formant extractor, Fig. 3: Linear predictive analysis section of 2211 FIG. 4 is a block diagram of the main part of the formant extraction device in the second embodiment of the present invention, and FIG. 5 is a block diagram of the main part of the formant extraction Hffi in the third embodiment of the present invention. It is a diagram. 11... Acoustic analysis unit, 12... Vowel discrimination unit, 13... Formant estimation coefficient storage unit, 1
4... Formant estimation section, 21... Linear prediction analysis section, 22... Higher order equation tree root section, 2
3... Formant a selection section, 41... Linear prediction analysis section, 42... Vowel discrimination section, 43...
... Formant estimation coefficient storage section, 44 ...
Formant estimation section, 511...Linear prediction analysis section, 512...Cepstrum calculation section, 52...
. . . Vowel discrimination section, 53 . . . Formant estimation coefficient storage section, 54 . . . Formant estimation section.

Claims (3)

[Claims] (1) an acoustic analysis unit that performs acoustic analysis on input vowel audio waveform data to calculate audio features representing a spectral envelope; and a vowel discrimination unit that performs vowel discrimination of the input vowel audio waveform data based on the audio features; Based on the voice feature data and formant frequency data of the corresponding plural vowels of multiple speakers in advance, a regression coefficient obtained by performing a multiple regression analysis to estimate the formant frequency from the voice feature amount for each vowel by gender and each vowel is calculated. The corresponding vowel,
a formant estimation coefficient storage unit that outputs a gender regression coefficient; and a formant estimation unit that calculates and outputs a formant frequency estimate as a regression line value from the voice feature amount, the corresponding vowel, and the gender regression coefficient. A formant extraction device comprising: (2) The formant extraction device according to claim (1), characterized in that an α parameter is used as the audio feature. (3) The formant extraction device according to claim (1), characterized in that LPC cepstral coefficients are used as voice features.