JPS62194300A - Pitch extraction system - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to speech analysis, and particularly to an improvement of a pitch extraction method suitable for extracting the pitch period of speech in real time.

[Conventional technology]

Pitch period information is extremely important in terms of sound quality in high-efficiency encoding, speech synthesis, etc. in which speech is decomposed and transmitted or stored, and a means for extracting it with high precision is indispensable. In addition, real-time processing is essential especially in transmission, and a high-speed pitch extraction algorithm with low processing amount is desired in order to reduce equipment costs.

The pitch frequency of voice is 70-5, considering men and women.
00H7 (period: 2 to 15 ms), and the extraction accuracy is preferably 0.1 ms or less, or at least 0.3 ms or less, from the viewpoint of the quality of encoded speech or synthesized speech. Conventionally, analog-digital (A/D
) Pitch synchronization with sufficient time resolution was extracted by simultaneously using the five converted signals.

In order to extract pitch synchronization, the autocorrelation coefficient of the speech waveform or prediction residual waveform is calculated for a delay of 2 to 15m5, and the pitch period is determined by the time delay value that gives the peak value of the correlation coefficient. is common. Now 8k
Considering the case of Hz sampling, the time delay per sample is 1.25 μs, so 2 to 15 m s
The delay corresponds to a delay of 16 to 120 sample points, and considering the reliability of the extracted autocorrelation coefficient, matching data of about 100 points is required even for the value with a delay of 120 samples, which is the largest delay. , the audio sample data is 220
The calculation amount for the autocorrelation coefficient with a lag of 16 to 120 points becomes very large.

By reducing the calculation amount of '/il for the pitch extraction mentioned above, a general-purpose signal processing microcomputer (DSP) can be used to
For example, Japanese Patent Application Laid-Open No. 57-82897 is a method that can be realized in a time period of 10 to 20 ms). This is the amount of data obtained by resampling the input audio to, for example, 1/4 and then calculating the correlation coefficient.

This method reduces the amount of calculations and performs parabolic interpolation around the peak value of the correlation coefficient to extract the pitch period while ensuring the necessary time resolution. Further, Japanese Patent Application Laid-Open No. 76891/1983 discloses a method in which low-order linear predictive analysis is performed during beam sampling, and pitch extraction is performed after removing the influence of formants. Further, JP-A-58-1140798 discloses a method of determining a guide index from the pitch period of the past few frames 11 and extracting pitch synchronization by taking into consideration the continuity of the pitch period.

[Problem that the invention seeks to solve]

The above-mentioned Japanese Patent Application Publication No. 57-8289 or Japanese Patent Application Publication No. 58-76891
If the technology disclosed in 2003 is applied to voice transmitted over a telephone line, sufficient performance may not always be obtained. This is because voice with a limited band (300 to 3400 Hz) is susceptible to the effects of harmonic components of the voice. In other words, the harmonic components thereof are emphasized relatively more than the pitch components, and - which is an integer of the true pitch period is more likely to be selected. Conversely, a period that is an integer multiple may be selected. These problems occur even at normal sampling rates, but when resampling is performed, the above problems increase due to the mismatch between the true pitch period and the sampling period.

On the other hand, since pitch synchronization is extracted independently for each frame 11, discontinuity is likely to occur. On the other hand, JP-A-58-114
The method of 0798 is effective in maintaining the continuity of pitch periods. However, since the jl function value for pitch period candidates is not evaluated, if there are many errors in the extracted pitch period, there is a possibility that the errors will propagate. It is necessary to select the pitch period again after extracting the pitch period. This means that the code has an additional delay of 80 ms, and the impact on speech quality cannot be ignored.

An object of the present invention is to provide a highly accurate pitch extraction method that requires less data and less processing, and essentially requires less encoding delay.

[Means for solving problems]

In order to achieve the above object, the present invention calculates a plurality of candidates from the pitch period extracted by the peak of the correlation coefficient of the sampled audio signal, and calculates the correlation value of the short section of the audio signal for each of them. Through evaluation, the most appropriate pitch period is selected from these candidates. Also, at this time, by weighting the correlation values based on the pitch periods extracted up to the previous frame, a stable pitch period with guaranteed continuity is selected.

[Effect]

Waveform 21 in FIG. 3 shows an example of a voice waveform. Furthermore, a section 31 indicates the frame t1 from which the pitch is extracted. The autocorrelation coefficient is calculated using the following equation for a waveform y obtained by resampling the waveform X+ at a ratio of 4:1, which is obtained by reducing and filtering the original waveform x1 (i-th sampled waveform).

Let the time delay for interpolating the vicinity of the maximum value of R(t) and giving the maximum value be T (having the resolution of the original sampling).

At this time, the pitch period candidates are T, nT, and T/n (n is an integer of 2 or more) within the pitch period search range. T/3. T/2, T, and 2T are separate sections (each g5
This shows the result of calculating the correlation value with the sections 33, 34, 35, 36) in Figure 3 using the following equation.

t:O Here, X is the amplitude of the sample data of the i-th audio waveform, and the beginning of the section 32 is conveniently set at i=Q.

M is the predetermined number of data, and j is the section 33, 34, 35, .
36 first data number (address), that is, T/3
．． T/2. T, 2T (however, an integer). According to Figure 3, r(T/2) has a value similar to r(T),
The correct pitch period can be determined to be T/2. Here, the formula (2
) is the original data, that is, the data before being resampled, so if the reference section (section 32 in Figure 3) is appropriately selected, stable judgments can be made with a relatively small amount of data. .

By the way, the pitch period candidate is generally T/n.

There are both nT, but if limited to either case,
It becomes easy to judge the pitch. Therefore, R(t) of formula (1)
By multiplying by an appropriate weight W(t) according to the order t, T that gives the maximum value of R(t) will extract only the correct pitch period or its integer multiple, and the formula (
r(j) in 2) only needs to be calculated for j=T/n (n≧1), and in this case, the pitch period should be the smallest j among those that satisfy r(j)”r(T). .

On the other hand, R(t) is calculated for each frame.

In rare cases, a value that is a non-integer multiple of the correct pitch is selected, and for such cases it is not possible to extract the correct pitch period. The above problem can be avoided by selecting a pitch period that is considered to be continuous and close to the pitch period extracted up to the previous frame 11. Specifically, the weight W (t) multiplied by R(t) may be made larger than the standard value only in L corresponding to the pitch period of the immediately previous frame and its vicinity. A similar operation can be applied to the case where the pitch period is selected by evaluating the value of r (T/n) (n is an integer of 1 or more) in equation (2). That is, r( for T/n close to the pitch period of the immediately previous frame)
What is necessary is to weight the correlation value of T/n).

The feature of this method is that the continuity of the pitch period is reflected in the correlation coefficient. By doing this, even if the pitch period extracted in the previous frame is incorrect, the correlation value corresponding to the correct pitch period in the current frame can be larger than the incorrectly weighted correlation value. It has a high level of accuracy, making it difficult for errors to propagate.

〔Example〕

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

FIG. 1 is a block diagram of a pitch extraction device using the present invention. In FIG. 1, audio data digitized with predetermined sampling synchronization is stored in a buffer memory 1. Here, the sampling period is 125 μ5 (8 kHz sampling) and the frame period is 20 m5. The buffer memory stores 40m5 worth of data centered on the current frame. Audio data X is read out from buffer memory 1, inputted to resampling section 2, and
A waveform yi resampled at kHz is output.

In the resampling unit 2, the audio data xi passes through a low-pass filter with a cutoff frequency of 500 Hz, and
have been thinned out.

The resampled audio data yi is input to the autocorrelation coefficient calculation unit 3, and the autocorrelation coefficient RD is calculated according to equation (1).
) is calculated. Here, the pitch period search range is 2 to 1
5 m F! and for 8 kHz sampling, t
win = 16 ~ tmax = 'L corresponds to 20 samples, but for a waveform resampled at 2 kHz corresponds to 4 to 30 samples. However, since pitch period candidates are extracted by parabolic interpolation, it is necessary to calculate R(t) for two extra samples, that is, for a time delay t=3 to 31.

R(t) is input to the pitch period candidate extraction unit 4, and first weighted.

R' (t)=R(t) ・W(t) (3
) W (t, ) is as shown in FIG. 5, for example. This is a kind of low-frequency emphasis, and has the effect of suppressing extraction of an integer part of the correct pitch period.

Next, the maximum value of R' (t) corresponding to t=4 to 30 is detected. Give the maximum value of R'(t).The order is t=t
o, the pitch period candidate T is determined by the following parabolic interpolation to obtain the time resolution (125
μS).

From the pitch period candidate extracting unit 4, the maximum value R of R' (t)
'(to) is output to the determination unit 6, and the pitch period candidate T is output to the partial correlation calculation unit 5, respectively.

In the partial correlation calculation unit 5, audio data X is read out from the buffer memory 1, and T/n≧τwin (5)
r(T/n) is calculated according to equation (2) for T/n. Here n is an integer greater than or equal to 1, and T/
n is a value expressed as an integer. Here, we will explain how to find the start address of the reference section, where soil=0 in equation (2) for convenience.

The purpose of the partial correlation calculation unit 5 is to calculate the correlation coefficient of equation (2) by T/
The purpose is to find it with good sensitivity for n. For this purpose, it is desirable to use the part with the highest periodicity in the frame as the reference. An example of how to find the reference interval is to first
is detected, the sum of the absolute values of the amplitudes is determined for M continuous audio data including that data, and k '' k o that gives this maximum value is set as the start address of the reference section.Equation (6) ) of a(k
) may be used instead of power.If the reference interval is determined in this way, 1
The number of data M in equation (2) is the minimum pitch period τwin
It turns out that about twice the amount is sufficient.

The address ko determined in this way is set again to i=O, and equation (2) is calculated. Here, when ko is in the latter half of the frame, equation (2) may be used instead.

From the partial correlation calculation unit 5, T/n and r (T/n
) is output to the determination section 6.

The determining unit 6 first determines whether the frame is voiced or unvoiced by threshold value determination of the output R'(To) from the pitch extraction candidate extracting unit 4. That is, R' (t,)≧0
1 (7) is voiced, and r(
T/n). Here, θl is a positive threshold. Otherwise, the frame is considered unvoiced, the pitch period is 7, and =O is output, and the processing of the frame is terminated.6 If the frame is voiced, the output of the partial correlation calculation unit is T/
using n and r(T/n).

r (T/n';: r (T) - 02 (8) The smallest one among T/n (n is an integer of 2 or more) that satisfies (8) is defined as the pitch period. However, θ2 is a positive threshold value. If the expression (
If T/n satisfying 8) does not exist, the pitch period is set to τ=T.

By outputting the pitch period 7, the processing of the frame Is is completed.

Next, a second embodiment of the present invention will be described using FIG. 2.

The difference from the first embodiment is that a weight control section 8 is added. This is for the purpose of more stable pitch extraction by using the pitch information up to Frame 11 immediately before the current frame.The current 0 weight control unit 8, when the pitch period of the immediately preceding frame is determined, The following processing is performed.

In the weight control unit 8, the pitch period τ2 of one frame and two frames before the current frame is stored, and 1 τ knee τ21≦Os
(9) When the first control parameter PL becomes PL=τ
1/4 (1o) When the second control parameter P2 is set as P2=τs (11) and does not satisfy equation (9), P1=O(1
0)' P z = O(11)' is set. Here, 08 is a positive threshold value, and represents the width at which the fluctuation in pitch period between two consecutive frames can be regarded as continuous.

The processing in this frame is the first one up to the autocorrelation calculation unit 3.
This is similar to the embodiment. In the pitch period candidate selection section 4, the value of the autocorrelation coefficient R(t) is partially corrected by the first control parameter pt supplied from the weight control section 8. That is, here Wl is usually 1 with a weight of 1 or more.
．． It is about 1 to 1.2. In addition, ΔP indicates the width of the order to be corrected, and is about 1 to 2. In Equation 1 (12), Wl is assumed to be constant, but it is also possible to form a mountain-shaped weight around t:=Pt. . By formula (12), the previous 2
If continuous pitch synchronization is extracted in a frame, a period close to it is likely to be selected.

On the other hand, in the determination section 6, the second control parameter Px supplied from the weight control section 8 is determined based on the correlation coefficient r' (T/n) (n is an integer of 1 or more) supplied from the partial correlation calculation section. Selective weighting is performed by That is, n≧
l T/n −P x I≦04 for 1
If (13) is satisfied, r (T/n) = r (T/n) ・vtx (
14). Number 0 is the fluctuation range of the pitch period that can be regarded as continuous, and is usually 0+ = θδ. Moreover, wz is a weight of 1 or more. When such weighting is performed, r max = wax ((']-/n) (15
)77 mouths, and instead of the determination using equation (8), for n≧1, r(T/n)≧r+wax −Oz' (16
), the smallest one of T/n (n≧1) that satisfies T/n (n≧1) may be used as the pitch period. 02' is a threshold value based on 02. At the time when the pitch period τ of the frame is determined, τ2
The values are updated as follows: = τl, τ1 = τ.

The simplest example of how to determine the control parameters Px and Pz in the weight control unit 8 has been shown. It goes without saying that there are various variations in the method of determining control parameters.

The processing of the first embodiment and the second embodiment described above requires a relatively small amount of computation and memory, and can be easily implemented using a general-purpose microprocessor or the like. When the pitch of speech transmitted via a telephone line was extracted in the second embodiment, the extraction error was reduced from about 25% to 5%.

〔Effect of the invention〕

According to the present invention, pitch period candidates can be extracted with a small amount of processing and accurate determination can be made in consideration of the continuity of pitch periods, so that pitch periods can be extracted more accurately.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a pitch extraction device according to a first embodiment of the present invention, FIG. 2 is a block diagram of a pitch extraction device according to a second embodiment of the present invention, and FIG. 3 is a diagram showing audio waveforms. FIG. 4 is a diagram showing the principle of the present invention, and FIG. 5 is a diagram showing a weighting function. 2...Resampling unit, 3...Autocorrelation coefficient calculation unit. 4... Pitch period candidate extraction section, 5... Partial correlation calculation section, 6... Judgment section, 8... Weight control section.

Claims (1)

[Claims] 1. In a pitch extraction method in which a digitized audio signal is divided into frames of a predetermined time length and a pitch period is extracted from the correlation function value of the audio signal within the frame, the pitch period is extracted from the correlation function value. A pitch period candidate is extracted, a time delay correlation value corresponding to synchronization of the pitch candidate is precisely calculated with respect to the audio signal in the frame, and a pitch period is extracted based on the correlation value. Pitch extraction method. 2. The pitch extraction method according to claim 1, wherein a predetermined weighting is applied to the correlation function value according to the order. 3. The pitch extraction method according to claim 2, wherein the weight is controlled based on the pitch period extracted up to the immediately previous frame. 4. The correlation function value of the audio signal within the frame is calculated by interpolating the correlation function value calculated after thinning the audio signal within the frame to a predetermined ratio. A pitch extraction method according to claim 1, 2, or 3.