JPH026080B2 - - Google Patents

Description

[Detailed Description of the Invention] Industrial Application Field The present invention divides a speech waveform into arbitrary sections, performs two nonlinear transformations on each section, calculates the cross-correlation coefficient of the two variables, and calculates the pitch. This invention relates to a pitch extraction device that performs extraction.

Configuration of conventional example and its problems The voiced part of the speech waveform has a periodic repeating waveform, and the change characteristics of the period (pitch) are as follows.
It is known that it is an important parameter in audio processing. In speech analysis and synthesis systems,
The extraction result of pitches extracted during analysis has a great influence on the quality of synthesized speech during synthesis.

Conventionally, as a method for extracting the pitch of an audio waveform, a method has been widely used in which an autocorrelation coefficient of an audio signal is calculated and extracted for each frame in which the audio signal is divided into a certain length of time. However, the above method is not suitable for real-time analysis of speech because it incorrectly extracts double-period or half-period components of the correct pitch as pitch, and requires a lot of calculation time due to the complexity of the calculation. . Furthermore, when considering the hardware configuration for performing real-time analysis, there are drawbacks such as the need for a high-speed arithmetic processing device for the above reasons, resulting in a large-scale device.

As a method to shorten the calculation time, the audio waveform x(n) is expressed as follows: y ₁ (n)=x(n)−C _L x(n)≧C _L =0 |x(n)|<C _L = x(n)+C _L x(n)≦-C _L ...(1) y ₂ (n)=1 x( _{n)≧C L =} 0 | )≦−C _L ...(2) However, C _L : A method has been proposed in which pitch is extracted by performing non-linear transformation using a threshold value and calculating the cross-correlation coefficient of the non-linearly transformed binary values. . (LAWRENCE R.
RABINER：On the Use of Autocorrelation
Analysis for Pitch Detection, IEEE Trans.
ASSP-25, No. 1, 1977) However, this method also has the drawback that components such as double periods and half periods of the correct pitch are erroneously extracted as pitches.

Purpose of the Invention The present invention was made to solve the conventional problems as described above.
It is an object of the present invention to provide a pitch extracting device which requires less total calculation processing amount than conventional ones and can obtain highly accurate pitches.

Structure of the Invention In order to achieve this object, the present invention non-linearly transforms an audio waveform divided into arbitrary sections using a threshold value to obtain y ₁ (n) and y ₂ (n), and converts these into In the process of finding the cross-correlation value P ₍ 〓 ₎ of the function and setting the pitch to the value of τ that maximizes P ₍ 〓 ₎ , the threshold value is divided by dividing the speech waveform into the first half and the second half. The maximum value of the two parts is determined separately, a new circuit is added to determine the maximum value from the two determined maximum values, and the threshold value is made variable for each arbitrary section, thereby reducing the amount of calculation, and This makes it possible to prevent incorrect extraction of pitches.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram for explaining an embodiment of the present invention.

A voice waveform is input to an A/D converter 1.
The A/D converter 1 samples the audio waveform at a predetermined frequency (normally, if the sound quality is comparable to that of a telephone voice, the sampling frequency is 8KHz, and 8 bits is sufficient for sampling the audio waveform, (This will be explained below using these numerical values.) A sampled audio waveform obtained by converting the audio waveform into a discrete time-series signal is output to the data buffer memory 2. The data buffer memory 2 temporarily stores the sampled audio waveform, and when the sampled audio waveform for the analysis frame period is stored, the first half of the sampled audio waveform is sent to the first half maximum value detection circuit 3. The second half is output to the second half maximum value detection circuit 4. Here, the data amounts of the sampled audio waveforms in the first half and the second half are equal. Furthermore, data buffer memory 2
outputs all of the sampled audio waveforms to the nonlinear conversion circuit 6 and the ternary classification circuit 7.

The first half maximum value detection circuit 3 determines the maximum value MAX1 of the sampled audio waveform, and the threshold value determination circuit 5
Output to. The second half maximum value detection circuit 4 determines the maximum value MAX2 of the second half of the sampled audio waveform,
It is output to the threshold value determining circuit 5.

The threshold value determination circuit 5 determines the threshold value C _L based on the maximum value MAX1 and the maximum value MAX2 using the following equations (3), (4), and (5).

IMAX=max (MAX1, MAX2) ………(3) IMAX1=min (MAX1, MAX2) ………(4) C _L = min (IMAX・0.6, IMAX1・0.8) ………(5) Here max ( , ) is the larger of the two,
min(,) means to select the smaller one.

The threshold determining circuit 5 applies the determined threshold C _L to a nonlinear conversion circuit 6 and a ternary classification circuit 7.
Output to.

Here, the threshold value determining method of the present invention will be explained using FIG. 2.

It is clear that pitch is T in analysis frame A. In order to detect the pitch T, it is necessary to set the threshold level below the peak value 201, that is, the threshold value 202. Next, in analysis frame B, the pitch is T1. However, if the threshold value 202 is adopted, the peak value 203 may exceed the threshold value 202, and a period twice the true pitch T1 may be taken during pitch extraction. In the threshold determination method of the present invention, the analysis frame B is divided into a first half _B1 and a second half _B2 , and the first half maximum value detection circuit 3 and the second half maximum value detection circuit 4 detect the first half. The maximum value is the wave height value 204, and the maximum value in the second half is the wave height value
205 is detected, and then the threshold value determination circuit 5 determines a value of 60% of the peak value 205 as the threshold value 206. As a result, true pitch T1 is obtained in pitch extraction.

Similarly, if the threshold value determination method of the present invention is used for analysis frame A, 80% of the peak value 201 will be determined as the threshold value (same level as the threshold value 202), so in pitch extraction. The true pitch T is obtained.

The nonlinear conversion circuit 6 nonlinearly converts the sampled audio waveform transferred from the buffer memory 2 and temporarily stored according to the equation (1) based on the threshold value C _L. FIG. 3 shows an example in which the sampled speech waveform is nonlinearly transformed by the nonlinear transformation circuit 6. In FIG.
FIG. 3a shows a sampled speech waveform, and FIG. 3b shows a sampled speech waveform after nonlinear transformation. Furthermore, the nonlinear conversion circuit 6 outputs the nonlinearly converted sampled audio waveform to the pitch calculation circuit 8.

The ternary classification circuit 7 classifies the sampled audio waveform transferred from the buffer memory 2 and temporarily stored into ternary values according to the equation (2) based on the threshold value C _L , and calculates the pitch. Output to calculation circuit 8. Fourth
The figure shows an example in which the sampled speech waveform is classified into three values by the three-value classification circuit 7 as shown in equation (2) above. Figure 4a
is the sampled speech waveform, and b is the signal after ternary classification.

The pitch calculation circuit 8 receives y ₁ (n), y ₂ (n) supplied via the nonlinear conversion circuit 6 and the ternary classification circuit 7.
Based on the signal, calculate the cross-correlation coefficient P ₍ 〓 ₎ as shown in equation (6) below.

P ₍ 〓 ₎ = _N- 〓 〓 ^N=1 y ₁ (i)・y ₂ (i+τ) ………(6) (i=1,……N) However, the number of sampled audio waveforms per frame is N pieces.

The pitch change range for normal adult men and women is 50Hz~
400Hz, and searching this range, the range of τ is τ=20 to 160. Let Pmax(τo) be the one that takes the maximum value among P ₍ 〓 ₎ obtained from the above equation (6), and then τo is taken as the pitch.

Effects of the Invention As explained above, the present invention changes the threshold value C _L every analysis frame period depending on the amplitude state of the audio waveform, and nonlinearly transforms the audio waveform using the value _. (n), y ₂ (n), y ₁ (n),
By calculating the cross-correlation coefficient of y ₂ (n),
It takes less calculation time and allows highly accurate pitch extraction.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining a pitch extraction device according to an embodiment of the present invention, FIG. 2 is an explanatory diagram for explaining a threshold value determination method, and FIG. FIG. 4 is a diagram showing the characteristics of the ternary classification circuit 7. DESCRIPTION OF SYMBOLS 1... A/D converter, 2... Data buffer memory, 3... First half maximum value detection circuit, 4... Second half maximum value detection circuit, 5... Threshold determination circuit, 6
...Nonlinear conversion circuit, 7...Thinary classification circuit, 8...
...Pitch calculation circuit.

Claims (1)

[Claims] 1. The speech waveform is divided into arbitrary sections and pitch extraction is performed for each section, and within each section, the speech waveform x(n) is calculated by the following formula y ₁ (n)= x(n)−C _L x(n)≧C _L =0 |x(n)| _{<C L =} x( _n )+C _L )= 1 x(n)≧C _L =0 |x(n)|<C _L =-1 x(n)≦-C _L
......(2) However, C _L : Nonlinear transformation is performed according to the threshold value, and y ₁ (n) and y ₂ (n) are calculated, and the mutual phase of y ₁ (n) and y ₂ (n) is When the pitch is determined by calculating the relational coefficient, the audio waveform within the section is divided into the first half and the second half, and the above threshold is determined for each section from the amplitude information of the audio waveform of the two parts. A pitch extraction device featuring: