JP3195700B2 - Voice analyzer - Google Patents

Abstract

PURPOSE:To prevent the deterioration in tone quality by remarkably reducing an error rate of pitch extraction from an audio signal and the error rate of discrimination of sound/soundless. CONSTITUTION:The minimum value and the nearest maximum value of a normalized mean amplitude difference function are detected, and the true minimum value is obtained from the minimum value by an interpolater 9, and a difference between the true minimum value and the nearest maximum value is obtained by an ALU 10 to be inputted to a pitch detector 11, and further, a guide pitch detected by the detector is obtained by the ALU 12 to be inputted to the pitch detector 11. In addition, the audio signal from an A/D converter 3, an extraction parameter from a spectal envelope parameter extractor 4 and the true minimum value and the nearest maximum value of the normalized mean amplitude difference function corresponding to the detected pitch from the pitch detector 11 are inputted to a sound/soundless discriminator 13.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech analyzer for automatically analyzing and extracting pitch and voiced / unvoiced sound discrimination parameters, which are basic parameters of sound source information, from a speech signal.

[0002]

2. Description of the Related Art Conventionally, typical examples of this type of speech analysis method include (a) an autocorrelation method, (b) a modified correlation method,
(C) SIFT algorithm, (d) Average amplitude difference function (AMDF: Average Magnitude D)
In any case, for automatic pitch extraction,
A major cause of sound quality degradation during speech synthesis is erroneous extraction of a pitch twice as long as the true fundamental period (double pitch) or half of the pitch (half pitch) or erroneous discrimination between voiced and unvoiced sounds. Had become.

[0003]

SUMMARY OF THE INVENTION In view of the above problems in speech analysis, the present invention significantly reduces the pitch extraction error rate and voiced / unvoiced discrimination error rate, thereby minimizing sound quality degradation. It is an object of the present invention to provide a voice analysis device which is prevented.

[0004]

According to the present invention, there are provided (1) an input terminal for an audio signal, a first low-pass filter, an analog-to-digital converter, a spectrum envelope parameter extractor, and a second low-pass. A pass filter, and an average amplitude difference function calculator for calculating a normalized average amplitude difference function of audio waveform data of a fixed time length (one frame) digitized at a predetermined sampling period through the filters. A pitch detector that detects the pitch of the fundamental frequency of the audio signal from the delay amount at which the normalized average amplitude difference function obtained in step 1 takes a minimum value. An average amplitude difference function minimum value / maximum value detector for detecting the minimum value and the nearest maximum value of the normalized average amplitude difference function obtained, and detection of three or more points including the minimum value detected by this detector Value And a interpolator to determine the true minimum value at a quadratic curve more approximate function which approximates,
The true minimum value obtained by the interpolator is input to the pitch detector, and the pitch (IP) of the previous frame detected by the pitch detector and output to the pitch output terminal is output.
k-1) and a guide pitch (GPk-1), and a guide pitch calculator for calculating a guide pitch (GPk) that approximates the pitch (IPk) of the frame currently being analyzed. ) Is input to the pitch detector, and the pitch detector detects a pitch (IPk) by giving different identification conditions between a pitch search range before and after the guide pitch (GPk) and other search ranges. The above problem is solved by providing a voice analysis device. (2) In the speech analyzer according to the above (1), an average amplitude difference function minimum value / maximum value difference calculator for calculating a difference between a true minimum value obtained by the interpolator and a nearest maximum value is provided. And the difference (DEL
TRAi) is input to the pitch detector, and the speech signal converted by the analog-to-digital converter, the spectrum envelope parameter (K1) of the parameter extracted by the spectrum envelope parameter extractor, and the least square error (E
p) and the difference (DELTRA) between the true minimum value and the most recent maximum value of the normalized average amplitude difference function corresponding to the pitch detected by the pitch detector, as a discrimination parameter. The above object is attained by providing a voice analysis device including a voiced / unvoiced sound discriminator for discriminating voiced / unvoiced sound under conditions.

[0005]

According to the present invention, a new function is found for the difference between the average amplitude difference functions of the residual signals, and a correction function having a coefficient corresponding to the pitch of the fundamental frequency and a guide pitch for giving an expected value of the pitch to be extracted are used in common. By doing so, the pitch extraction error rate and the discrimination error rate between voiced and unvoiced sounds are significantly reduced.

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to a block diagram of a speech analyzer shown in FIG. The audio signal of a fixed time length (one frame) applied to the audio signal input terminal 1 is sampled by the first low-pass filter 2 at a sampling frequency (for example, 8 KH).
z) at least a half or more frequency component (for example, 3.
(4 kHz or more), and is digitized by an analog-to-digital converter (A / D converter) 3 at a sampling frequency.

Next, after the digitized signal is cut out by the data window in the spectral envelope parameter extractor 4, an autocorrelation function is obtained.

Further, if a partial correlation analyzer is used to extract the spectral envelope parameters, the spectral envelope parameters K1, K2... Kp and the least square error Ep can be obtained by, for example, the following cutting equation.

Kn = Wn-1 / Un-1 n = 1,... P (p: analysis order)

[0010]

(Equation 1)

(Equation 2)

[0011]

(Equation 3)

Ep = Up / Vo Spectral envelope parameter K thus obtained
1,... Kp and the digitized audio signal sequence are added to the analysis filter 5 or directly to the second low-pass filter 6 as indicated by a dashed line arrow according to the characteristics of the sound source.

When a partial correlation lattice type filter expressed by the following equation is used as the analysis filter 5, a residual signal sequence EDj is obtained by the following equation.

EFi = EFi-1−Ki · EBBi−1 i = 1,... P EBi = EBBi−1−Ki · EFi−1 i = 1,. EBBo = NDj-1 EDj = EFp, j = 1 ... 3, SN (SN: number of cut-out audio signals) The obtained residual signal sequence or the audio signal sequence is removed or attenuated from components having a frequency substantially equal to or higher than the maximum pitch frequency. Through a second low-pass filter 6 to filter a high-frequency component from a residual signal sequence or an audio signal sequence, EA (j = 1,
.., SN).

Further, EAj is converted to an average amplitude difference function (AMDF)
It is input to the arithmetic unit 7 to obtain a normalized average amplitude difference function sequence RAj by the following equation.

[0016]

(Equation 4)

In the present invention, attention is paid to the point that if the minimum RAj is simply obtained and j is set as the pitch, the probability of an extraction error for obtaining the double pitch, triple pitch, and half pitch increases. Further, the obtained normalized average amplitude difference function sequence RAj is input to a detector 8 for detecting the minimum value / maximum value of the AMDF, and the minimum values RAMNi and RAMN of RAj are detected.
A value APi (a value of a pitch selection candidate) of j giving i and a local maximum value RAMAXi of RAMNi are obtained.

Although the maximum value RAMAXi exists on a relatively gentle curve, the minimum value RAMNi exists on a sharp peak curve, so that an accurate value cannot be obtained in the sampling period.

Then, the minimum value RAMNi is passed through the AMDF minimum value interpolator 9 to determine the true minimum value.

For such an interpolation method, for example, it is sufficient to approximate a local minimum value with a quadratic curve (y = ax + bx + c) and set its vertex to a true local minimum value RAMINi.

Next, the true minimum value and the nearest maximum value are expressed by AM
The difference is input to the DF minimum value / maximum value difference calculator 10 and the difference DELTRAi between the minimum value / maximum value is calculated by the following equation.

DELTARAi = RAMAXi-RAMIN
i Further, this difference DELTARi and the true minimum value obtained by the interpolator 9 are input to the pitch detector 11, and the pitch IP is detected by means described later, and the detected pitch IP is sent to the pitch output terminal 14. It is output and input to the guide pitch calculator 12.

On the other hand, the voice signal converted by the A / D converter 3 and the parameters extracted by the spectrum envelope parameter extractor 4 are input to the voiced / unvoiced sound discriminator 13 to discriminate the voiced / unvoiced sound. Are output to the voiced / unvoiced discrimination parameter output terminal 15 and also to the guide pitch calculator 12. It should be noted that the input to the guide pitch calculator 12 (the one-dot chain line arrow) should be practical, but a certain effect can be obtained without it.

The guide pitch calculator 12 calculates the guide pitch G in the k-th speech frame analysis.
The following calculation for Pk is performed.

GPk = (1-GR) .GPk-1 + GR.I
Pk-1 where GR is a suitable constant of 0 <GR <1. IPk-1 is the extracted pitch of the (k-1) th frame obtained immediately before.

At this time, preferably, a continuous unvoiced frame (including a non-voice frame) is converted to a voiced frame.
A voiced sound frame after several voiced sound frames continues is defined as a fully voiced sound frame, and a voiced sound frame up to the fully voiced sound frame is referred to as a transition frame to a voiced sound. At this point, the average pitch between transition frames to voiced sounds is set as the new guide pitch.

By inputting the determined guide pitch GPk of the k-th frame to the pitch detector 11, the pitch IPk is selected from APi (pitch selection candidate value) in the pitch detector 11 in the following procedure. Is selected.

(1) For example, the guide pitch GPk <32
Then, the variable GPC is set to 1.375. GPk <64
Then, the variable GPC is set to 1.25. If GPk ≧ 64, the variable GPC is set to 1.1875.

(2) For example, the minimum expected pitch GPMIN
= GPk / GPC, for example, the maximum expected pitch GPM
AX = GPk · GPC (the pitch search area during this period is the expected pitch existence area to be extracted).

(3) j = 1.

(4) After a lapse of about one frame,
RAI, RAJ
Weight variable is introduced. The current analysis frame (kth) is a completely voiced frame and APi ≧ GPMI
N and if APi≤GPMAX, weight variable RAI
= RAC, APj ≧ GPMIN and APj ≦ GPMA
If X, the weight variable RAJ = RAC. RAC
Is, for example, RAC = 0.34375. Conversely, if the current analysis frame is an unvoiced sound frame or a transition frame to a voiced sound, the variable RAI = RAJ = 0.

(5) When the normalized average amplitude difference functions have the same magnitude, weight variables WAI and WAJ are introduced due to the fact that a smaller pitch is more likely to be a correct pitch. It is assumed that the weight variable WAI = APi · RAWGT. It is assumed that the weight variable WAJ = APj · RAWGT. The constant RAWGT is, for example, RAWGT = 1/512.

(6) RAMINi + WAI-RAI> R
If AMINj + WAJ-RAJ, the process proceeds to (8). Conversely, RAMINi + WAI-RAI ≦ RAMINj
If it is + WAJ-RAJ, the process proceeds to (7).

(7) IPk = APi, DELTA = DE
Let LTRAi, j = i.

(8) If the candidate for APi still exists,
Return to (4) and repeat the process.

(9) When the selection operation is completed for all APi, the final IPk becomes the pitch to be obtained and is output to the pitch output terminal 14.

In this case, as shown in FIG. 2, for example, in the conventional AMDF method, since RAMN2 is the minimum value, AP2 is extracted as a pitch. In the present invention, the true minimum is obtained by performing interpolation. Value RAMIN1, RA
MIN2 is found, RAMIN1 is found to be the true minimum, AP1 can be extracted as pitch,
AP2 is prevented from being erroneously extracted and becoming double pitch. Further, in a completely voiced sound frame, as shown in FIG. 3, for example, even if RAMINi> RAMINj due to the introduction of the guide pitch and the weighting variable, the correct pitch Aj is used instead of the double pitch APj.
Extract Pi as the pitch IPk of the k-th frame.

The spectrum envelope parameter K 1, the least square error Ep, and the difference DELTRA between the minimum value and the maximum value thus obtained are input to the voiced / unvoiced sound discriminator 13. Is determined based on the following criteria. Note that this reference value is an example. (A) K1 ≥ 0.9 and DELTA * 0.15625
Then it is a voiced sound. (B) K1 ≧ 0.7, Ep ≦ 0.5 and DELTA *
If it is 0.193375, it is a voiced sound. (C) K1 ≧ 0.4 and Ep ≦ 0.3 and DERTRA ≧
If it is 0.23125, it is a voiced sound. (D) K1 ≧ 0 and Ep ≦ 0.2 and DERTRA ≧ 0.
26875 is a voiced sound. (E) Ep <0.7 and DERTRA ≧ 0.30625
And if Ep <DELTRA + 0.15, it is a voiced sound. (F) Ep ≧ 0.7 and DERTRA ≧ 0.34375
If Ep <DELTRA + 0.3, it is a voiced sound. (G) When Ep is extremely small (for example, Ep <0.00)
In 1), even if (a) to (f) are satisfied, the voice is unvoiced. (H) When the audio signal level is extremely low,
Even if (f) is satisfied, the sound is unvoiced. (G) Those that do not meet the above conditions shall be unvoiced. The discrimination result is output from the voiced / unvoiced discrimination parameter output terminal 15.

In the unvoiced sound section, as shown in FIG. 4, for example, the magnitude of the normalized average amplitude difference function is often determined to be a voiced sound by the conventional method. Then, by introducing the idea of the difference DELTRA between the true minimum value and the maximum value of the normalized average amplitude difference function, in the case of the example in FIG. 4, since DELTRA is small in the example of FIG. This frame can be determined to be unvoiced.

Further, as shown in FIG. 5, even if RAMIN is large enough to be judged as unvoiced sound in the conventional judgment method, if DELTRA is sufficiently large, the combination of other conditions will cause the present invention to fail. In the invention, it can be determined that the voiced sound is present.

[0041]

As described above, according to the first aspect of the present invention,
The voice analysis device detects the minimum value and the latest maximum value of the normalized average amplitude difference function with a detector, obtains a true minimum value from the detected minimum value with an interpolator, and outputs the true minimum value to a pitch detector. Input, and furthermore, a guide pitch which gives an expected area of the pitch detected by this pitch detector is obtained by a guide pitch calculator and input to the pitch detector, and a weight variable is introduced into the pitch detector. The error rate of extraction can be significantly reduced, and the deterioration of sound quality can be prevented as much as possible.

Further, according to the second aspect of the present invention,
The true minimum value and the nearest local maximum of the normalized average amplitude difference function corresponding to the voice signal converted by the analog-to-digital converter, the parameters extracted by the spectrum envelope parameter extractor, and the pitch detected by the pitch detector Since each value of the difference from the value is input to the voiced / unvoiced sound discriminator, the discrimination error rate of the voiced / unvoiced sound can be significantly reduced, and the deterioration of the sound quality can be prevented as much as possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a characteristic diagram of a normalized average amplitude difference function and delay showing an example of the effect of interpolation of a minimum value.

FIG. 3 is a characteristic diagram of a normalized average amplitude difference function and a delay illustrating an example of an effect of a weight variable in a completely voiced sound frame.

FIG. 4 is a characteristic diagram of a normalized average amplitude difference function and a delay illustrating an example in which an unvoiced sound is easily erroneously determined to be a voiced sound.

FIG. 5 is a characteristic diagram of a normalized average amplitude difference function and delay showing an example in which a voiced sound is likely to be erroneously determined to be unvoiced.

[Explanation of symbols]

 Reference Signs List 1 audio signal input terminal 2 reduction pass filter 3 analog-to-digital converter 4 spectrum envelope parameter extractor 5 analysis filter 6 low-pass filter 7 AMDF calculator 8 AMDF minimum / maximum value detector 9 AMDF minimum value interpolator 10 AMDF minimum Value / maximum value difference calculator 11 pitch detector 12 guide pitch calculator 13 voiced / unvoiced sound discriminator

Claims (2)

(57) [Claims] 1. An audio signal input terminal, a first low-pass filter, an analog-to-digital converter, a spectrum envelope parameter extractor, a second low-pass filter, and a predetermined sampling period through these. An average amplitude difference function calculator for calculating a normalized average amplitude difference function of the digitized audio waveform data having a fixed time length (one frame), wherein the normalized average amplitude difference function obtained by the calculator is A pitch detector for detecting a pitch of a fundamental frequency of the audio signal from a delay amount having a minimum value, wherein a minimum value of the normalized average amplitude difference function obtained by the average amplitude difference function calculator is provided. Mean amplitude difference function for detecting the minimum value and the local maximum value, a minimum value / maximum value detector, and an approximation function of a quadratic curve or more approximated from three or more detection values including the minimum value detected by this detector And an interpolator for obtaining a true minimum value.The true minimum value obtained by the interpolator is input to the pitch detector, and detected by the pitch detector and output to a pitch output terminal. A guide pitch calculator for calculating a guide pitch (GPk) that approximates the pitch (IPk) of the frame currently being analyzed from the pitch (IPk-1) of the previous frame and the guide pitch (GPk-1);
The obtained guide pitch (GPk) is inputted to the pitch detector, and the pitch detector receives the guide pitch (GPk).
k) A voice analysis device that detects a pitch (IPk) by giving different identification conditions between the preceding and following pitch search ranges and other search ranges. 2. The speech analyzer according to claim 1, wherein an average amplitude difference function minimum value / maximum value difference calculator for calculating a difference between a true minimum value obtained by the interpolator and a nearest maximum value is provided. And the difference (DELTR
Ai) is input to the pitch detector, and the speech signal converted by the analog-to-digital converter, the spectrum envelope parameter (K1) of the parameter extracted by the spectrum envelope parameter extractor, and the least square error (Ep)
And a difference (DELTRA) between the true minimum value and the latest maximum value of the normalized average amplitude difference function corresponding to the pitch detected by the pitch detector, as a determination parameter, A voice / unvoiced sound discriminator for discriminating voiced / unvoiced sounds.