JPH0683390A - Speech analyzing method - Google Patents

Abstract

PURPOSE:To facilitate arithmetic operation and make a voiced/voiceless decision with high likelihood by performing specific arithmetic operation for analyzing the tendency of variation in autocorrelation function and generating a correction value, and making the voiced/voiceless decision. CONSTITUTION:In a step S8, the specific arithmetic operation for analyzing the tendency of variation in autocorrelation function Ri is performed and the correction value is generated according to the arithmetic operation result. For the purpose, the number LPCT (Local Peak Count) of extreme value of the autocorrelation function Ri is counted. In a step S9, it is checked whether or not the number LPCT of extreme value found in the step S8 is less than 24 (namely, LPCT-24<0) and it is judged which of two correction expressions is applied. Thus, an arithmetic operation expression obtained by combining the correction expression and a decision expression together is executed, the value MAXR of the ratio of a voiceless area is corrected, and voiced/voiceless decision arithmetic operation based upon the correction value MAXR' is carried out to obtain a voiced/voiceless decision value UV.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice analysis method, and more particularly to a method for determining voiced sound and unvoiced sound, and more particularly to a method for easily and quickly determining voiced sound and unvoiced sound. The method for determining voiced sound and unvoiced sound according to the present invention can be used, for example, in the field of musical sound control technology for electronic musical instruments and the like to perform musical sound control according to the result of determination of voiced sound and unvoiced sound.

[0002]

2. Description of the Related Art As one of methods for determining voiced sound and unvoiced sound, there is a method using a modified autocorrelation function re (n). This method is roughly divided into the following four steps.
Step 1: By linear prediction, a residual waveform signal of the speech waveform signal to be analyzed is obtained. Step 2: Obtain an autocorrelation function of this residual waveform signal (this is called a modified autocorrelation function). Step 3: The ratio re of the initial value re (0) of this modified autocorrelation function and the maximum extremum re (T) thereafter.
Calculate (T) / re (0). Step 4: The voiced / unvoiced voice signal is determined from the obtained ratio value. As an example, if the value of the above ratio is 0.18 or less, it is determined as unvoiced sound, and the voicedness determination value V = 0 and the unvoicedness determination value U = 1. If the value of the ratio is 0.25 or more, it is determined as voiced sound, and the voicedness determination value V = 1 and the unvoicedness determination value U = 0.
And If the value of the ratio is between 0.18 and 0.25, the voicedness determination value V and the unvoicedness determination value U are appropriately determined according to the values (provided that U + V = 1).

[0003]

As described above, the voiced / unvoiced decision method using the modified autocorrelation function is not easily influenced by the formant component in the speech waveform signal, and therefore accurate decision can be expected. However, on the other hand, there is a drawback that the configuration of the arithmetic unit and the program becomes complicated because a complicated arithmetic operation for obtaining the residual waveform is required. Also, since the calculation time will be extra,
It took time from the input of the voice waveform signal to the output of the voiced / unvoiced determination result, and it was difficult to perform the analysis in real time. Therefore, input an arbitrary voice waveform signal and
It is not suitable for performing musical tone control in real time according to the result of unvoiced judgment. On the other hand, it has been considered that the autocorrelation function of the speech waveform signal to be analyzed is directly obtained and the voiced / unvoiced decision is made based on this autocorrelation function, but the problem is that the decision accuracy that can be put to practical use alone cannot be obtained. There was a point. The present invention has been made in view of the above points, and an object of the present invention is to provide a voice analysis method suitable for real-time analysis, which enables relatively accurate voiced / unvoiced determination with a simple configuration. To do.

[0004]

A voice analysis method according to the present invention comprises a first step of inputting a voice waveform signal, a second step of calculating an autocorrelation function of the input voice waveform signal, and To analyze the tendency of the change in the autocorrelation function obtained in the second step and the third step of calculating the ratio between the initial value of the autocorrelation function obtained in the second step and the maximum extremum thereafter. Is performed and a correction value is generated according to the calculation result, and the ratio value obtained in the third step is corrected according to the correction value generated in the fourth step. And a fifth step of performing voiced / unvoiced determination based on the corrected ratio value.

[0005]

In the second step, the autocorrelation function of the input voice waveform signal is directly obtained. Therefore, the calculation for obtaining the modified autocorrelation function is omitted, and the calculation is considerably simplified. In the third step, the ratio between the initial value of the autocorrelation function obtained in the second step and the maximum extremum thereafter is calculated. Generally, in the autocorrelation function,
The maximum extreme value after the initial value appears at the position corresponding to the pitch period of the voice waveform signal. Therefore, the periodicity of the voice waveform signal is known from the value of the ratio between the initial value of the autocorrelation function and the maximum extreme value thereafter, and based on this, it is possible to determine voiced sound and unvoiced sound. However, in the autocorrelation function of the voice waveform signal, components other than the fundamental component remain, so that accurate voiced / unvoiced determination cannot be performed only by the value of the ratio. In particular, since a random extreme value due to a noise component in unvoiced sound is generated at a relatively large level, it is extremely difficult to make a voiced / unvoiced determination only by the value of the ratio. Thus, the value of the ratio obtained in the third step is
It is insufficient.

The present invention is characterized by including the fourth and fifth steps for correcting the insufficient value of the ratio. Especially, in the autocorrelation function of the speech waveform signal, the influence of the noise component remains, so that the autocorrelation function of the unvoiced sound has a larger change in the autocorrelation function value than the autocorrelation function of the voiced sound. (Or intense)
It focuses on that. That is, the tendency of the function waveform to change in the autocorrelation function of the speech waveform signal is
It focuses on the relationship with unvoicedness. Therefore, in the fourth step, a predetermined calculation for analyzing the tendency of the change of the autocorrelation function obtained in the second step is performed, and the correction value is generated according to the calculation result. I am trying. And in the fifth step,
The value of the ratio is corrected according to this correction value, and voiced / unvoiced determination is performed based on this. Accordingly, for example, if the autocorrelation function value tends to change significantly, the unvoiced degree is high. Therefore, if the value of the ratio is corrected in the direction of increasing the unvoiced degree, the voice with relatively high accuracy can be obtained. It will be possible to make a silent judgment. Moreover, since the calculation for obtaining the modified autocorrelation function is omitted, the calculation configuration is simple and the calculation time is shortened accordingly. Therefore, it is also suitable for real-time analysis.

There are various conceivable analytical calculation methods in the fourth step. One embodiment of the analytical operation may consist in counting the number of extreme values in the autocorrelation function determined in the second step. If the number of extreme values is large, it means that the tendency of the waveform change of the autocorrelation function is strong (or frequent), and the unvoicedness is high. In this case, both the maximum value and the minimum value may be counted as the extreme values to be counted, or either the maximum value or the minimum value may be counted.

One embodiment of the method of counting the number of extreme values in the above is, in each function value of the autocorrelation function,
The difference between adjacent function values may be obtained, and counting may be performed when the sign of the difference changes. When the sign of the difference changes from positive to negative, it means that the maximum value has passed, and when the sign of the difference changes from negative to positive, it means that the minimum value has passed, so that number is counted. Therefore, the number of extreme values can be counted.

Another embodiment of the analytical operation in the fourth step may consist of accumulating the absolute values of the differences between the maximum and minimum values in the autocorrelation function obtained in the second step. The larger the accumulated absolute value of the difference between the maximum value and the minimum value, the greater the difference between the maximum value and the minimum value, and the greater the number of the extreme values. Therefore, the amplitude change of the autocorrelation function is large and changes. It means that the tendency is large (or intense), and the degree of unvoicedness is high.

Yet another embodiment of the analytical operation in the fourth step consists in accumulating the absolute values of the differences between the function values for each predetermined sample in the autocorrelation function obtained in the second step. Good. The larger the accumulated absolute value of the difference between the function values for each predetermined sample is, the larger the difference between the function values for each predetermined sample is. Therefore, the change in the autocorrelation function is steeper and the change tendency is large (or severe). It means that there is a high degree of silence. In this case, the absolute value of the difference between the function values of adjacent samples in the autocorrelation function may be accumulated, or the absolute value of the difference between the function values of a suitable number of samples may be accumulated. Good.

An embodiment of the correction in the fifth step may be that the correction is performed according to a different correction formula depending on whether or not the correction value is a predetermined value or more. This is effective for performing different types of correction depending on whether the correction value indicates the unvoiced degree or the voiced degree, and performing a more accurate voiced / unvoiced determination.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the accompanying drawings. FIG. 1 is a flow chart showing a processing procedure of an embodiment of a voice analysis method according to the present invention. First, an analog voice waveform signal is externally fetched by a microphone or the like (step S1). Next, the captured analog voice waveform signal is low-pass filtered (step S2). Next, the low-pass filtered analog voice waveform signal is converted into a digital signal (step S3).

In the next step S4, the digitally converted voice waveform signal sample data is stored in the memory. In this case, the waveform sample data Y (m + i) for one frame required for one analysis may be stored in the memory. Here, m is a number indicating an arbitrary initial sample, and indicates the initial sample number of the frame to be analyzed. i is 1
A variable that indicates the relative sample number within the frame, i =
0, 1, 2, ..., N. For example, 1 frame = N +
When 1 sample = 128 samples, i = 0, 1, 2,
..., 127. In the following, for convenience, m = 0
The waveform sample data for one frame stored in the memory is indicated by Y0 to Y127. In the next step S5, the waveform sample data Y for one frame stored in step S4
0 to Y127 are read out and the autocorrelation function Ri thereof is calculated.
An example of the calculation formula of the autocorrelation function is as follows.

[0014]

[Equation 1] Here, i in Ri is a variable indicating the order of the autocorrelation function, and i = 0, 1, 2, ..., 127. N is
As described above, it is the final sample number in one frame. n is a variable that changes from i to N, and specifies the sample number of the waveform sample data Yn (Y0 to Y127). Therefore, in the above formula 1, when n = i, Yn
= Yi, Yn-i = Y0, and when n = N = 127, Yn = Y127 and Yn-i = Y127-i. According to the calculation formula of the autocorrelation function as shown in the above mathematical expression 1, a total of N + 1 = corresponds to each value (0, 1, 2, ..., 127) of i.
An autocorrelation function value Ri of 128 samples is obtained. The obtained autocorrelation function value Ri is stored in the memory.

In step S6, the initial value R0 and the maximum extremum Rk excluding the initial value R0 are extracted from the autocorrelation function value Ri obtained in the previous step (for R0 and Rk, for example, FIG. 6 (e)). See). In step S7, the ratio Rk / R0 between the maximum extreme value Rk and the initial value R0 is obtained. In the following, this ratio Rk / R0 is represented by MAXR. It is generally known that the interval between the initial value R0 and the maximum extreme value Rk corresponds to the pitch of the original speech waveform signal. Therefore,
This ratio Rk / R0 = MAXR is effective for knowing the voiced degree of the original speech waveform signal. However, since the maximum extremum Rk occurs at random positions in unvoiced sound, this ratio Rk / R0 = MAXR is not sufficient to know the unvoiced degree of the original speech waveform signal.

As a result of many measurements, the ratio Rk /
It was found that the relationship between the value of R0 = MAXR and the voicedness / unvoicedness shows a tendency as shown in FIG. That is, not only does the voiced sound show a high value, but also the unvoiced sound shows a relatively high value. The transitional sound between voiced sound and unvoiced sound has a relatively low value. Therefore, the value of this ratio M
It is difficult to make an accurate voiced / unvoiced determination only with AXR. In addition, as a result of such actual measurement, the value M of the above ratio
It has been found that by applying a hypothetical relational expression such as the following Expression 2 according to AXR, it can be used as a standard for determining the voiced / unvoiced level. Since the numerical values 0.4 and 0.8 in the following formula 2 are merely empirical values, when generalizing the hypothetical formula of the following formula 2, these numerical values 0.4 and 0.8 are used. Of course, it should be replaced with an appropriate constant according to the conditions.

[0017]

## EQU00002 ## If MAXR.ltoreq.0.4 holds, unvoiced sound MAXR.gtoreq.0.8 holds, if voiced sound 0.4 <MAXR <0.8, transient state As described above, since the ratio value MAXR itself is inaccurate, an accurate determination cannot be made only with the relational expression. Therefore, in order to compensate the inaccuracy of the ratio value MAXR, the following steps S8 to S15 are performed,
The value MAXR of this ratio is corrected, and based on this, the final voiced / unvoiced judgment value UV is obtained with relatively high accuracy.

In step S8, a predetermined calculation for analyzing the change tendency of the autocorrelation function Ri obtained in step S5 is performed, and a correction value is generated according to the calculation result.
As an example of the analytical calculation for that, step S8 of FIG.
Then, the number of extreme values LPCT in the autocorrelation function Ri obtained in the previous step S5 (the meaning of this symbol is Local PeakCo
I want to be understood as unt).
In this case, both the maximum value and the minimum value may be counted as the extreme values to be counted, or either one of the maximum value and the minimum value may be counted. As is usually defined, the maximum means an upward peak (peak) of the function waveform, and the minimum means a downward peak (valley) of the function waveform. Therefore, if the number of extreme values, that is, the number of peaks, the frequency of vibration of the function waveform is high, which means that the tendency of the change of the autocorrelation function is large,
Therefore, it means that the degree of unvoicedness is high.

As an example, the predetermined sampling frequency 1
The original speech waveform data sampled under 0 kHz is obtained by auto-correlation function Ri at 1 frame = 128 samples as described above, and a number of actual measurements are performed by the method of counting both the maximum value and the minimum value. It was found that the above-mentioned extreme number LPCT shows a tendency as shown in FIG. That is, it can be understood that the extreme value number LPCT shows a large value in unvoiced sound. Referring to FIG. 3, by appropriately correcting the value MAXR of the ratio in the unvoiced region,
It can be understood that if the function of the corrected ratio value shows a monotonically increasing tendency, accurate voiced / unvoiced determination can be performed based on this. Also, FIG.
, Since the extreme value LPCT shows a large value in the unvoiced region, it can be predicted that the ratio value MAXR, which is unstable in the unvoiced region, can be appropriately corrected by using this. . Therefore, the number of extreme values L
Properly correct the ratio value MAXR using PCT,
Based on this, it is proposed to make an accurate voiced / unvoiced decision as follows.

Various correction equations for such correction can be set. As a result of many measurements and trial and error, as an example, we have introduced that it is relatively effective to introduce the following correction formula and judgment formula. However, it goes without saying that the present invention is not limited to this. [MAXR correction formula] As a result of a number of actual measurements under the above sampling conditions and conditions in which both the maximum value and the minimum value are counted, as a guideline under the conditions,
It has been found that if the extreme value LPCT is smaller than "24", no correction is necessary. If the extremum number LPCT is higher, it is likely to be unvoiced sound or a transient region. Then, the ratio value MAXR of the unvoiced sound or the sound in the transient region is
It has been inferred that it is relatively effective to correct Eq. In addition, MAXR 'shows the value of the ratio after correction.

[0021]

(A) MAXR correction formula when LPCT is 24 or more: MAXR '= MAXR-√ (LPCT-24) / 32 (b) MAXR correction formula when LPCT is smaller than 24: MAXR' = MAXR (that is, without correction) Since the numerical values 24 and 32 in the above mathematical expression 3 are merely empirical values under the above conditions, in generalizing these numerical values, these numerical values 24 and 32 are Of course, it should be replaced with an appropriate constant according to the conditions.

[UV Judgment Formula] The voiced / unvoiced judgment value UV to be finally obtained is as follows by applying the ratio value MAXR ′ corrected by the above Equation 3 to the relational expression of the above Equation 2. It is calculated according to the relational expression of Equation 4.

[Equation 4] UV = (MAXR'-0.4) /0.4 (However, if MAXR'-0.4 <0, UV = 0, and if MAXR'-0.4> 0.4 Is set to UV = 1.) As can be understood from the above mathematical expression 4, if the ratio value MAXR ′ corrected by the mathematical expression 3 is 0.4 or less, the voiced / unvoiced judgment value UV is 0, and the unvoiced sound is unvoiced. Is shown. If the ratio value MAXR 'corrected by the above equation 3 is 0.8 or more, the voiced / unvoiced judgment value UV is 1,
Indicates voiced sound. Also, 0.4 <MAXR '<
When 0.8, voiced / unvoiced judgment value UV is 0 <UV <1.
Is in the voiced / unvoiced transitional area, and indicates the voiced degree or unvoiced degree according to the decimal value.

Returning to FIG. 1, in step S9, it is checked whether or not the number of extreme values LPCT obtained in the previous step S8 is smaller than 24 (that is, whether LPCT-24 <0 is satisfied).
This is to determine which of the correction equations a and b in the above Equation 3 should be applied. If YES is determined in step S9, the process proceeds to step S10. Step S1
At 0, the uncorrected ratio value MAXR is used to perform an operation of UV = (MAXR-0.4) /0.4. This corresponds to an arithmetic expression that is a combination of the correction expression b of the above-mentioned Expression 3 and the determination expression of the above-mentioned Expression 4.

If NO in step S9, the process proceeds to step S11. In step S11, the value MAXR of the ratio is corrected by the correction equation a of the above-mentioned equation 3, and UV = [{MAXR-√ (LPCT-24) / 32}-
[0.4] /0.4 is performed. This corresponds to an arithmetic expression that is a combination of the correction expression a of the above-mentioned Expression 3 and the judgment expression of the above-mentioned Expression 4. In this way, the correction equation of the above-mentioned equation 3 and the equation 4 are obtained by step S10 or S11.
An arithmetic expression that is a combination of the above judgment expressions is executed, and the ratio value MA
Along with the correction of XR, the voiced / unvoiced determination calculation based on this correction value MAXR 'is performed, and the voiced / unvoiced determination value UV is obtained.

In the subsequent steps S12 to S15, the processing according to the proviso condition of the above-mentioned equation 4 is performed. That is, when the voiced / unvoiced determination value UV obtained in the previous step S10 or S11 is larger than 1 (that is, MAXR'-0.4.
> 0.4), YES is determined in step S12,
Go to step 14 and set the voiced / unvoiced judgment value UV to 1. When the voiced / unvoiced determination value UV obtained in the previous step S10 or S11 is smaller than 0 (that is, MAX
R'-0.4 <0), YES is determined in step S14, the process proceeds to step 15, and the voiced / unvoiced determination value UV is set to 0.
Set to. From the above, accurate voiced / unvoiced judgment value UV
Can be obtained.

Although only the analysis for one frame has been described above, it goes without saying that the analysis for a plurality of frames is also possible. In that case, when dividing the waveform sample data into frames, the initial sample number m of each frame may be arbitrarily designated and each frame may be arbitrarily set. Further, in step S4, the digital voice waveform sample data over the number of samples of one frame or more are sequentially stored in the memory, the digital voice waveform sample data of a desired frame is read from the memory, and the processes of steps S5 to S15 are performed. Good.

Next, a specific example of the extreme value number counting process in step S8 of FIG. 1 will be described with reference to FIG.
In FIG. 2, for each function value of the autocorrelation function Ri obtained in step S5 of FIG. 1, a difference between adjacent function values is obtained, and when the sign of the difference changes, a maximum value (upward peak) or It is determined that the minimum value (downward peak) has been passed, and the number is counted. First, in step S20, the difference R between the initial value R0 and the next value R1 of the autocorrelation function Ri obtained in step S5 of FIG.
1-R0 is obtained, and this is set in the register d1. At the same time, the calculation variable j designating the order i of the autocorrelation function Ri is set to 1, and the count value of the extreme value number LPCT is set to 0. In step S21, the calculation variables j and j + 1
The difference between the autocorrelation function values Rj and Rj + 1 of two adjacent orders designated by is calculated and set in the register d2.

In the next step S22, the positive / negative of the difference value stored in each of the registers d1 and d2 is checked to determine whether the positive / negative sign has changed. That is, d
It is determined whether 1 ≧ 0 and d2 <0 holds, or whether d1 <0 and d2 ≧ 0 holds. d1 ≧ 0 and d
When 2 <0 holds, it means that the difference between the adjacent function values has changed from positive to negative, that is, has passed the maximum value. When d1 <0 and d2 ≧ 0 are satisfied, it means that the difference between the adjacent function values changes from negative to positive, that is, the minimum value is passed. If either of the determination conditions is satisfied, the process proceeds to step S23, and the number of extreme values LP
The count value of CT is incremented by 1, and then the process goes to step S24. If neither determination condition is satisfied, the process jumps to step S24 and LPCT is not counted.

In step S24, the difference value stored in the register d2 is transferred to the register d1 and the calculated variable j
Is increased by 1. In the next step S25, it is checked whether the calculation variable j has become 128, that is, whether the collation has been completed for all the orders i = 0, 1, 2, ... 127 of the autocorrelation function Ri. If NO, the process returns to step S21, and for the increased order j, j + 1, the same step S as described above is performed.
The processing from 21 to S24 is repeated. When the collation is completed for all orders i of the autocorrelation function Ri, step S
25 becomes YES, and the processing of FIG. 2 is terminated. At this time, the content of LPCT indicates the total count number of the maximum value and the minimum value in the autocorrelation function Ri. In addition, when it is desired to count only one of the maximum value and the minimum value, the determination condition in step S22 may be set to only one.

Next, an actual measurement example will be described. Figure 5
An example of a voice waveform when a male voice pronounces “sha” is shown. FIG. 6 shows autocorrelation functions in some frames obtained based on the speech waveform of FIG. In this example, the sampling frequency of the voice waveform is 10 kHz, and the number of samples in one frame is 128 (hence the time length of one frame is 12.8 msec). Figure 5
Is 1917 msec after 1817 msec from the start of measurement
Waveforms up to the latter are shown. This part generally starts with an unvoiced "s", goes through a transient state, and then goes to a voiced "a".
Since it is a part that shows the waveform change up to and including, and seems to be suitable for three kinds of analysis of unvoiced, transient, and voiced, it was selected as an actual measurement example.

FIG. 6A shows the autocorrelation function Ri of the speech waveform for one frame (for convenience, the first frame: F1) from 1810 msec. (B) shows the autocorrelation function Ri of the speech waveform for one frame (second frame: called F2) from 1826 msec. (C) shows the autocorrelation function Ri of the speech waveform for 1 frame (3rd frame: F3) from 1834 msec. (D) shows the autocorrelation function Ri of the speech waveform for 1 frame (fourth frame: F4) from 1842 msec. (E) shows the autocorrelation function Ri of the speech waveform from 1858 msec to one frame (fifth frame: F5). (F) shows the autocorrelation function Ri of the speech waveform for 1 frame (sixth frame: F6) from 1874 msec. Each frame F1 to F6 shown in FIGS.
The ratio Rk / R0 = MAXR of the maximum extremum Rk and the initial value R0 in each autocorrelation function, the number of extrema LPCT calculated as described above for each of these autocorrelation functions, and these UV judgment result obtained by applying each formula
Is shown in the table below.

[0032]

[Table 1]

FIG. 7 shows a voice message "ch
i) ”is an example of a voice waveform when pronounced. Figure 8
Shows the autocorrelation function in several frames obtained based on the speech waveform of FIG. 7. Also in this example, the sampling frequency of the voice waveform is 10 kHz, the number of samples in one frame is 128 (the time length of one frame is 12.8
msec). FIG. 7 shows the waveform from 691 msec to 793 msec after the start of measurement. This part is
In general, it is a part that shows the waveform change from unvoiced sound "c" to transient voiced sound "i", and is suitable for three types of analysis: unvoiced, transient, and voiced. Since it seems that there are some, it was selected as an actual measurement example.

FIG. 8 (a) shows the autocorrelation function Ri of the speech waveform for one frame (first frame: F1 for convenience) from 688 msec. (B) shows the autocorrelation function Ri of the speech waveform for one frame (second frame: F2) from 712 msec. (C) shows the autocorrelation function Ri of the speech waveform for one frame (third frame: F3) from 720 msec. (D) is an autocorrelation function R of the voice waveform for one frame (fourth frame: F4) from 724 msec.
indicates i. (E) shows the autocorrelation function Ri of the speech waveform for one frame (fifth frame: F5) from 728 msec. (F) shows the autocorrelation function Ri of the speech waveform for one frame (sixth frame: F6) from 760 msec. Ratio Rk / R0 of maximum extreme value Rk and initial value R0 in the autocorrelation function for each frame F1 to F6 shown in FIGS.
The following table shows MAXR, the number of extreme values LPCT calculated as described above for each of these autocorrelation functions, and the determination result UV obtained by applying these to each of the equations (3) and (4).

[0035]

[Table 2]

It can be seen from Tables 1 and 2 above that the value MAX of the ratio Rk / R0 before correction according to the invention is
If R is applied as it is and the voiced / unvoiced determination is performed according to the above-mentioned mathematical expression 2, it should have been erroneously determined to be a transient state in an apparent unvoiced sound portion (for example, F1 and F3 in Table 1). And F1, F2, F in Table 2
4) That is, all of them are correctly determined as unvoiced sound by the correction according to the extreme value LPCT.

Next, the processing procedure of another embodiment of the voice analysis method according to the present invention will be described with reference to FIG. 9, steps S1 to S7 and S12 to S15 are the same as steps S1 to S7 and S12 to S15 having the same numbers in FIG. 1, and steps S16 to S19 are replaced by steps S8 to S11 in FIG. . After obtaining the above-mentioned ratio Rk / R0 = MAXR in step S7, step S16
Is processed. In step S16, a predetermined calculation for analyzing the tendency of the change in the autocorrelation function Ri obtained in step S5 is performed, and a correction value is generated according to the calculation result. This is different from step S8 of 1. In this step S16, the absolute value of the difference between the maximum value and the minimum value in the autocorrelation function Ri obtained in the previous step S5 is accumulated, and the accumulated value LP
DSUM (The meaning of this symbol is Local Peak Differencia
l Be understood as Sum. ) Is output. The absolute value of the difference between the maximum value and the minimum value indicates the interval between the upward peak (peak) and the downward peak (valley) in the function waveform, and a large value indicates that the amplitude change of the function waveform is large. means. The larger the cumulative value of the absolute value of the difference between the maximum value and the minimum value, the greater the opening of the maximum value and the minimum value, and the more the number of extreme values. Therefore, the amplitude change of the autocorrelation function is large, It means that the tendency of change is large or severe, and that the degree of unvoicedness is high.

As an example, the predetermined sampling frequency 1
The original speech waveform data sampled under 0 kHz was subjected to a large number of actual measurement as a result of obtaining the autocorrelation function Ri of 1 frame = 128 samples as described above.
The local peak difference accumulated value LPDSUM is shown in FIG.
It was found that the following tendency was exhibited. That is, it can be understood that the accumulated value LPDSUM shows a large value in unvoiced sound. When analyzing, in order to eliminate the analysis noise due to the difference in the volume of the sampled voice,
It is necessary to normalize the accumulated value LPDSUM.
In the following example, the normalized accumulated value LP is obtained by dividing the simple accumulated value by the initial value R0 of the autocorrelation function Ri.
I am trying to find DSUM. Therefore, the numerical value of the local peak difference accumulated value LPDSUM described here is a value normalized as described above.

Referring to FIG. 11, since the local peak difference accumulated value LPDSUM has a large value in the unvoiced sound area, the use of this value can appropriately correct the unstable value MAXR of the ratio in the unvoiced sound area. You can predict what you can do. Therefore, it is proposed as follows that the value MAXR of the ratio is appropriately corrected using the local peak difference accumulated value LPDSUM, and the accurate voiced / unvoiced determination is performed based on this. The value for that LPDS
Various correction formulas can be set according to UM. As a result of many actual measurements and trial and error, it was inferred that it is relatively effective to introduce the following correction formula according to LPDSUM as an example. However, it goes without saying that the present invention is not limited to this. Under the above sampling conditions, as one guide, it was found that if the accumulated value LPDSUM is smaller than "10", no correction is necessary. Accumulated value LPD
If the SUM is higher, it is likely to be unvoiced or transient. Therefore, it has been inferred that it is relatively effective to correct the ratio value MAXR of the unvoiced sound or the sound in the transient region by a correction formula such as the following Expression 5. In addition, MA
XR 'shows the value of the ratio after correction.

[0040]

(5) (a) MAX when LPDSUM is 10 or more
Correction formula of R: MAXR '= MAXR- (LPDSUM-10) / 64 (b) MAXR when LPDSUM is smaller than 10
Correction formula: MAXR '= MAXR (that is, without correction) Since the numerical values 10 and 64 in the above mathematical expression 5 are merely empirical values under the above conditions, when generalizing this, Of course, the numerical values 10 and 64 of should be replaced with appropriate constants according to the conditions. Voiced / unvoiced judgment value UV to be finally obtained
Is the ratio value MAXR ′ corrected by the above equation 5,
Similar to the above, it is obtained by applying to the relational expression of the equation (4).

Returning to FIG. 9, in step S17, is the normalized local peak difference accumulated value LPDSUM obtained in the previous step S16 smaller than 10 (that is, LPD)?
SUM-10 <0 is satisfied). This is to determine which of the correction equations a and b in the equation 5 should be applied. If YES is determined in the step S17, the process proceeds to a step S18. In step S18,
Using the uncorrected ratio value MAXR, the calculation UV = (MAXR-0.4) /0.4 is performed. This corresponds to an arithmetic expression that is a combination of the correction equation b of the above equation 5 and the determination equation of the above equation 4.

If NO in step S17,
Go to step S19. In step S19, the value MAXR of the ratio is corrected by the correction equation a of the above-mentioned equation 5, and UV = [{MAXR- (LPDSUM-10) / 64}
-0.4] /0.4 is performed. This corresponds to an arithmetic expression that is a combination of the correction expression a of the above expression 5 and the determination expression of the above expression 4. In this way, the correction equation of the above equation 5 and the equation 4 are obtained by step S18 or S19.
An arithmetic expression that is a combination of the above judgment expressions is executed, and the ratio value MA
Along with the correction of XR, the voiced / unvoiced determination calculation based on this correction value MAXR 'is performed, and the voiced / unvoiced determination value UV is obtained. Further, by the processing of the following steps S12 to S15, when the value UV is larger than 1 or smaller than 0, the processing of setting 1 or 0, respectively, is performed as described above.

Next, a specific example of the processing for obtaining the local peak difference accumulated value LPDSUM in step S16 of FIG. 9 will be described with reference to FIG. In FIG. 10, in each function value of the autocorrelation function Ri obtained in step S5 of FIG. 9, the difference between adjacent function values is obtained, and when the sign of the difference changes, the maximum value or the minimum value is passed. Then, the difference between the adjacent maximum value and the minimum value is obtained, and the absolute value of the difference is accumulated. First, step S3
0, the autocorrelation function Ri calculated in step S5 of FIG.
The difference R1-R0 between the initial value R0 and the next value R1 is calculated and set in the register d. Next, it is checked whether the value of the register d is smaller than 0 (step S31). If YES, the sign flag SF1 is set to "-1" to indicate that it is negative (step S32). If NO, the sign flag SF1 is set to "1" to indicate that it is positive (step S33).

In step S34, the calculation variable j designating the order i of the autocorrelation function Ri is set to 1, and the local peak number n is reset to 0. Step S3
In 5, the difference between two adjacent autocorrelation function values Rj and Rj + 1 specified by the calculation variables j and j + 1 is calculated,
This is set in the register d. Next, it is checked whether the difference value stored in the register d is negative (step S3
6) If YES, the sign flag SF2 is set to "-1" to indicate that it is negative (step S37). NO
If so, the sign flag SF2 is set to "1" to indicate that it is positive (step S38).

In step S39, the values of the first sign flag SF1 and the second sign flag SF2 are compared with each other, and it is checked whether they do not match. The first sign flag SF1 includes
The sign of the difference between the preceding two adjacent function values Rj-1 and Rj is set, and the sign of the difference between the following two adjacent function values Rj and Rj + 1 is set in the second sign flag SF2. It is set. Therefore, when the sign of the difference between the adjacent function values changes, that is, when the maximum value or the minimum value is passed, both flags SF1 and SF2 do not match, and step S39 is YES. If YES in step S39, the flow advances to step S40 to store the autocorrelation function value Rj of the order designated by the variable j in the local peak value register LP (n) corresponding to the current value of the local peak number n. Next step S41
Then, the value of the local peak number n is incremented by 1.

If NO in step S39 or after step S41, the process proceeds to step S42, the value of the second sine flag SF2 is shifted to the first sine flag SF1, and the operation variable j is incremented by 1. In step S43, it is checked whether the operation variable j has become 128, that is, whether the collation has been completed for all orders i = 0, 1, 2, ... 127 of the autocorrelation function Ri. If NO, step S
Returning to step 35, the processes of steps S35 to S42 similar to the above are repeated for the increased orders j and j + 1. When the collation is completed for all orders i of the autocorrelation function Ri, step S43 becomes YES, and step S44
go to. At this time, the local peak number n indicates the total number of peaks (maximum value and minimum value) in the autocorrelation function Ri, and n local peak value registers LP
The value of each peak is stored in each of (0) to LP (n-1).

In step S44, the local peak difference accumulated value LPDSUM is reset to zero. Step S4
At 5, it is checked whether the local peak number n is 1 or less. If NO, the process goes to step S46 to set the calculation variable j to 1. In the next step S47, the j-1th local peak value register LP (j-1) and the jth local peak value register LP (j) are determined according to the calculation variable j.
The peak value (maximum value and minimum value) of is read, the absolute value of the difference | LP (j) -LP (j-1) | is calculated, and this is calculated by LPDSU.
Add to the current value of M.

Next, in step S48, the calculation variable j is set to 1
To increase. In step S49, it is checked whether or not the calculation variable j matches the local peak number n, that is, whether or not the calculation in step S47 has been performed for all the peaks. NO
If so, the process returns to step S47 and the increased variables j, j
With respect to +1, the same calculation in step S47 as described above is performed. When the calculation of step S47 is completed for all the peaks, step S49 becomes YES, and step S5
Go to 0. In this way, the absolute values of the differences between the adjacent maximum values and minimum values are successively accumulated for all the maximum values and the minimum values in the autocorrelation function Ri, and the accumulated value LPDS
UM is required. In step S50, the accumulated value LPDSUM obtained as described above is divided by the initial value R0 of the autocorrelation function Ri to obtain the normalized local peak difference accumulated value LPDSUM.

Next, an example of the results of analysis according to the second embodiment described above with respect to the actually measured examples shown in FIGS. 7 and 8 will be shown. The ratio Rk / R0 = MAXR of the maximum extreme value Rk and the initial value R0 in the autocorrelation function for each frame F1 to F6 shown in a to f of FIG.
And the normalized local peak difference accumulated value LPDSUM calculated as described above for each of these autocorrelation functions, and the determination result UV obtained by applying these to each of the equations 5 and 4, It looks like the following table.

[0050]

[Table 3]

Next, a processing procedure of still another embodiment of the voice analysis method according to the present invention will be described with reference to FIG.
In FIG. 12, steps S1 to S7 and S12 to S15
Are steps S1 to S7 and S12 to S1 having the same numbers in FIG.
This is the same process as 5 and the above-mentioned ratio Rk / R0 in step S7.
After obtaining = MAXR, the process of step S51 is performed.

In step S51, step S5
A predetermined calculation for analyzing the tendency of the change of the autocorrelation function Ri obtained in step S1 is performed, and a correction value is generated according to the calculation result. The method of analysis calculation for that purpose is step S8 and FIG. This is different from step S16. In this step S51, the absolute value of the difference between adjacent sample values in the autocorrelation function Ri obtained in the previous step S5 is accumulated, and the accumulated value ADSUM (the meaning of this symbol is
It should be understood as Absolute Differencial Sum). The larger the difference between the adjacent sample values in the autocorrelation function Ri, the steeper the waveform change at that portion. Therefore, the accumulated value ADSUM of the absolute values of the differences between the adjacent sample values in the autocorrelation function Ri
The larger is, the steeper the tendency of the waveform change in the whole autocorrelation function is, which means that the change tendency is larger or more severe, which means that the unvoiced degree is higher.

As an example, the predetermined sampling frequency 1
The original speech waveform data sampled under 0 kHz was subjected to a large number of actual measurement as a result of obtaining the autocorrelation function Ri of 1 frame = 128 samples as described above.
It has been found that the absolute difference accumulated value ADSUM exhibits a tendency as shown in FIG. That is, the accumulated value A
It can be understood that DSUM shows the largest value in unvoiced sound, sharply decreases in the transient part, and slightly increases in voiced sound. In the analysis, it is necessary to normalize the accumulated value ADSUM in the same manner as described above in order to eliminate the analysis noise due to the difference in the volume of the sampled voice. In the following example, the normalized accumulated value ADSUM is obtained by dividing the simple accumulated value by the initial value R0 of the autocorrelation function Ri. Therefore, the absolute difference accumulation value ADSUM described here
The numerical value of is a value normalized as described above.

Referring to FIG. 13, since the accumulated value ADSUM shows a large value in the unvoiced area, if this is used, the ratio value MAXR which is unstable in the unvoiced area is used.
Can be expected to be corrected appropriately. Therefore, using the accumulated value ADSUM, the value of the ratio M
It is proposed as follows that the AXR is appropriately corrected and an accurate voiced / unvoiced decision is made based on the AXR. Various correction formulas corresponding to the accumulated value ADSUM for that purpose can be set. As a result of numerous measurements and trial and error, as an example, AD
We have introduced that it is relatively effective to introduce the following correction formula and judgment formula according to SUM. However, it goes without saying that the present invention is not limited to this.

Under the above sampling conditions, the normalized accumulated value ADSUM is "1" as one guide.
If it is smaller than 2 ”, it turns out that no correction is necessary. If the accumulated value ADSUM is more than that, there is a high possibility that it is in the unvoiced sound or the transient region. Therefore, it was inferred that it is relatively effective to correct the value MAXR of the ratio of the unvoiced sound or the sound in the transient region by a correction formula such as the following Expression 6. In addition, MAXR 'shows the value of the ratio after correction.

[0056]

[Formula 6] (a) MAXR when ADSUM is 12 or more
Correction formula: MAXR '= MAXR- (ADSUM-12) / 64 (b) Correction formula of MAXR when ADSUM is smaller than 12: MAXR' = MAXR (that is, no correction) Since 64 and 64 are merely empirical values under the above conditions, it is needless to say that when generalizing and expressing these, these numerical values 12 and 64 should be replaced with appropriate constants according to the conditions. Is. Voiced / unvoiced judgment value UV to be finally obtained
Is the ratio value MAXR ′ corrected by the above equation 2,
Similar to the above, it can be obtained by applying to the relational expression of the equation (4).

Returning to FIG. 12, in step S52, is the normalized absolute difference accumulation value ADSUM obtained in the previous step S51 smaller than 12 (that is, ADS)?
UM-12 <0 is established). This is to determine which of the correction equations a and b in the equation 6 should be applied. If YES is determined in step S52, the process proceeds to step S53. In step S53, UV = (MAXR-0.4) /0.4 is calculated using the uncorrected ratio value MAXR. This corresponds to an arithmetic expression that is a combination of the correction expression b of the above-mentioned Expression 6 and the judgment expression of the above-mentioned Expression 4.

If NO in step S52,
Go to step S54. In step S54, the value MAXR of the ratio is corrected by the correction equation a of the above-mentioned equation 6, and UV = [{MAXR- (ADSUM-12) / 64}-
[0.4] /0.4 is performed. This corresponds to an arithmetic expression that is a combination of the correction expression a of the above equation 6 and the determination expression of the above equation 4.

As described above, in step S53 or S54, an arithmetic expression that combines the correction equation of the above equation 6 and the determination equation of the above equation 4 is executed to correct the ratio value MAXR, and at the same time, the correction value MAXR. A voiced / unvoiced determination calculation is performed based on 'to obtain a voiced / unvoiced determination value UV. Also,
In the subsequent steps S12 to S15, when the value UV is larger than 1 or smaller than 0, the processing of setting 1 or 0 is performed, as described above.

Next, an example of the result of analysis according to the third embodiment described above with respect to the actual measurement examples of FIGS. 7 and 8 will be shown. The ratio Rk / R0 = MAXR of the maximum extreme value Rk and the initial value R0 in the autocorrelation function for each frame F1 to F6 shown in a to f of FIG.
And the normalized absolute difference accumulation value ADSUM calculated as described above for each of these autocorrelation functions, and the determination result UV obtained by applying these to each of the equations 6 and 4, It looks like a table.

[0061]

[Table 4]

In the above description, in obtaining the absolute difference accumulated value ADSUM, the absolute value of the difference between adjacent sample values in the autocorrelation function Ri is accumulated, but the invention is not limited to this. You may make it accumulate | store the absolute value of the difference of the autocorrelation function value for every predetermined sample number (order), such as every 4 samples.

Next, an embodiment of an electronic musical instrument integrally equipped with a sound source system and a voice analysis system for carrying out the method according to each of the above embodiments will be described with reference to FIG. Figure 1
4, various processes are executed under the control of a microcomputer including a central processing unit (CPU) 10, a read only memory (ROM) 11 and a random access memory (RAM) 12. In ROM11,
A program for realizing each processing procedure as shown in the above embodiment is stored. Microphone 13
The analog voice signal taken in from the outside is given to the A / D converter 15 via the amplifier 14, sampled according to a predetermined sampling clock, and converted into a digital voice waveform signal. This digital voice waveform signal is stored in the RAM 12. The digital signal processor (DSP) 16 executes a programmed calculation formula at high speed, and the DSP 16 performs complicated calculation such as calculation of the autocorrelation function Ri. With the above configuration, each processing procedure for voiced / unvoiced analysis as shown in the above embodiment is executed under the control of the microcomputer. Voiced / unvoiced judgment value U obtained as a result of analysis
V is appropriately stored in a register in the CPU 10, RAM 12, etc., and is used for tone control.

The panel switch section 17 includes various switches for setting, selecting, and controlling the tone color, volume, pitch, effect, etc. of a musical tone, and a switch for instructing to take in an audio signal from the outside or Voiced as a result of analysis /
Various switches related to the voice analysis process, such as a switch for instructing to use the unvoiced judgment value UV for tone control, may be included. The panel switch unit 17 is connected to the bus of the microcomputer via the interface 18. A keyboard section 19 having a plurality of keys for designating a scale note to be generated is also connected to a bus of a microcomputer via an interface 20.

A digital tone signal corresponding to the scale tone designated by the keyboard section 19 is generated from the tone signal generation circuit 21 and converted into an analog signal by the D / A converter 22.
Sound is produced acoustically from the sound system 23. As the tone signal generating method in the tone signal generating circuit 21, any known method such as PCM waveform memory method, FM method, AM method may be used.

Voiced / unvoiced judgment value UV obtained as a result of analysis
May be used for musical tone control in any form. For example, when it is used to control the tone color of a tone signal generated by the tone signal generation circuit 21, it controls the characteristics of the filter, controls the selection of the waveform read from the waveform memory, and controls the parameters of various interpolation calculations. The voiced / unvoiced determination value UV may be used for this purpose. The voiced / unvoiced determination value UV may be used in various other forms, but will not be described in detail here.

The control mode is, for example, the keyboard section 19
While sounding a musical sound specified by, etc., an appropriate voice is input by the microphone 13 and its analysis is performed in real time, and the musical sound being sounded is determined according to the voiced / unvoiced judgment value UV obtained as a result of the analysis. It is possible to control in real time. In that case, analyze only one appropriate frame,
The musical sound may be controlled according to the voiced / unvoiced determination value UV obtained as a result. Alternatively, an appropriate plurality of frames may be analyzed in real time, and the musical sound may be variably controlled in time according to the voiced / unvoiced determination value UV that may change according to each analysis result. .

The voice analysis system for carrying out the method according to each of the above embodiments may be configured as a single voice controller. FIG. 15 shows such a voice controller 2.
4 shows an example. In FIG. 15, the voice controller 24 includes a microphone 25 for taking in a voice signal from the outside, an amplifier 26, an A / D converter 27, a voice analysis device 28, and a MIDI converter 29. The voice analysis device 28 is a device for performing the method according to each of the above-described embodiments, and may be configured by a microcomputer equipped with the voice analysis program as described above and a DSP, or by a discrete circuit. Good. The MIDI converter 29 sends the voiced / unvoiced judgment value UV obtained as a result of the analysis and other necessary data to MIDI.
Convert to format and output. Voice / unvoiced judgment value UV in MIDI format and other necessary data are
It is given to an external electronic musical instrument 30 and used for controlling a musical sound generated in the electronic musical instrument 30.

In each of the above embodiments, the method of inputting the voice signal to be analyzed is not limited to the example of sampling and input by the microphone, but the digital voice waveform data may be input via the data transmission line. Alternatively, the digital voice waveform data already recorded in the memory may be appropriately read and input.

The method of analyzing and correcting the tendency of the change in the autocorrelation function of the voice waveform signal is not limited to the one shown in the above embodiment, and all modes that can be understood and predicted from the present specification and the drawings are applicable. Including. That is, the value of the ratio between the initial value R0 of the autocorrelation function directly obtained from the speech waveform signal and the maximum extreme value Rk is particularly unstable in the unvoiced sound region, and the idea that improvement of judgment in the unvoiced sound region is desired. Since the tendency of the change in the autocorrelation function changes more drastically for unvoiced sounds, the intensity of this change is quantified from some point of view, and the value of the ratio is changed according to this numerical value so that unvoiced sound regions can be particularly distinguished. Due to the idea of correction, accurate voiced / unvoiced determination can be expected in the present invention. Therefore, from such a viewpoint, it suffices to perform an operation for measuring / analyzing any factor that can quantify the degree of change in the autocorrelation function. All such embodiments are within the scope of this invention.

[0071]

As described above, according to the present invention, the autocorrelation function of the speech waveform signal is directly obtained, and based on this, the initial value of the function for rough voiced / unvoiced determination and the maximum extremum thereafter are calculated. The value of the ratio is obtained, further, a predetermined calculation for analyzing the tendency of the change of the autocorrelation function is performed, a correction value is generated according to the calculation result, and the value of the ratio is calculated according to the correction value. Since the correction is performed and the voiced / unvoiced determination is performed based on this, the calculation can be simplified as compared with the case where the modified autocorrelation function is obtained, and the corrected voiced / unvoiced determination can be performed with high accuracy. become able to,
It has an excellent effect. Therefore, it is most suitable when the voice signal is analyzed in real time and the musical tone control is performed in real time according to the analysis result.

[Brief description of drawings]

FIG. 1 is a flowchart showing a processing procedure of an embodiment of a voice analysis method according to the present invention.

FIG. 2 is a flowchart showing one specific example of the extreme value number counting processing step in FIG.

FIG. 3 is a graph showing an example of the relationship between the value of the ratio of the initial value of the autocorrelation function of the speech waveform signal and the maximum extreme value and the voicedness / unvoicedness in a rough tendency.

FIG. 4 is a graph showing an example of the relationship between the number of extreme values in the autocorrelation function of a speech waveform signal and the voicedness / unvoicedness in a rough tendency.

FIG. 5 is a diagram showing an example of a voice waveform in the case of pronouncing “sha” by a male voice.

6 is a functional waveform diagram showing an autocorrelation function in several frames obtained based on the speech waveform of FIG.

FIG. 7 is a view showing an example of a voice waveform when a female voice pronounces “chi”.

8 is a functional waveform diagram showing an autocorrelation function in several frames obtained based on the speech waveform of FIG. 7.

FIG. 9 is a flowchart showing a processing procedure of another embodiment of the voice analysis method according to the present invention.

10 is a flowchart showing a specific example of processing steps for obtaining a local peak difference accumulated value in FIG.

FIG. 11 is a graph showing an example of the relationship between the accumulated local peak difference value and the voiced / unvoiced degree in the autocorrelation function of the speech waveform signal, with a rough tendency.

FIG. 12 is a flowchart showing a processing procedure of still another embodiment of the voice analysis method according to the present invention.

FIG. 13 is a graph showing an example of the relationship between the absolute difference accumulated value and the voiced level / unvoiced level in the autocorrelation function of a voice waveform signal, with a rough tendency.

FIG. 14 is a block diagram showing an embodiment of an electronic musical instrument integrally equipped with a sound source system and a sound analysis system for implementing the sound analysis method according to the present invention.

FIG. 15 is a block diagram showing an embodiment in which a voice analysis system for implementing the voice analysis method according to the present invention is configured as a single voice controller.

[Explanation of symbols]

10 ... Central processing unit (CPU), 11 ... Read only memory (ROM), 12 ... Random access memory (RAM), 13 ... Microphone, 16 ... Digital signal processor (DSP), 17 ... Panel switch section,
Reference numeral 19 indicates a keyboard portion, 21 indicates a tone signal generating circuit, 24 indicates a voice controller, 28 indicates a voice analysis device.

Claims (1)

[Claims] 1. A first step of inputting a speech waveform signal, and a second step of calculating an autocorrelation function of the input speech waveform signal.
And a third step of calculating a ratio between the initial value of the autocorrelation function obtained in the second step and the maximum extremum thereafter, and a change of the autocorrelation function obtained in the second step. A fourth step of performing a predetermined calculation for analyzing the tendency and generating a correction value according to the calculation result, and the third step according to the correction value generated in the fourth step
5. A voice analysis method comprising: a fifth step of correcting the ratio value obtained in the step, and performing voiced / unvoiced determination based on the corrected ratio value.