JP3190231B2 - Apparatus and method for extracting pitch period of voiced sound signal - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device and method for extracting a pitch period capable of rapidly and easily extracting the pitch period by the processing of a time area. SOLUTION: A voiced sound collected by a microphone 11 is supplied to the pith period extraction device constituted of a microcomputer 13 through a line amplifier 12. After a positive peak and a negative extreme after the positive peak are detected from a voiced sound signal, the peak inverting the polarity of the positive peak is added to the negative extreme, and clamping processing is performed. Detection and differential processing are performed after the clamping processing is performed, and the pitch period is detected surely by measuring a pulse period.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus and a method for extracting a pitch period of a voice signal, and more particularly to a method and an apparatus for extracting a pitch period of a voiced sound signal which can be processed in a time domain.

[0002]

2. Description of the Related Art In recent years, practical use of a speech recognition device or a speech synthesis device has been promoted. However, extraction of a so-called pitch period from speech or a musical tone is important as one of basic techniques. Conventionally, various methods have been proposed as a method of extracting a pitch period, but they are roughly classified into a waveform processing method, a spectrum processing method, and a correlation processing method.

[0003] Waveform processing methods include a parallel processing method, a data reduction method and a zero-crossing method, spectrum processing methods include a cepstrum analysis method, and correlation processing methods such as an autocorrelation method or a linear prediction residual signal method. is there.

[0004]

However, most of the pitch period extraction methods proposed in the above proposals are extraction methods in the frequency domain. In order to convert a voice signal into the frequency domain by FFT or the like, a high-speed and large amount of calculation is required. Is required. Therefore, it is unavoidable that the pitch period extracting device becomes large and expensive.

SUMMARY OF THE INVENTION The present invention has been made in consideration of the above problems, and provides a pitch cycle extracting apparatus and a pitch cycle method capable of extracting a pitch cycle in real time and easily in time domain processing. With the goal.

[0006]

According to a first aspect of the present invention, there is provided a pitch period extracting apparatus, comprising: a voiced sound capturing means for capturing a voiced sound signal as time-series data; and a predetermined processing for the voiced sound signal captured by the voiced sound capturing means. Peak detection means for detecting a peak present in one area of the threshold value, and extreme value detection for detecting an extreme value existing in the other area of a predetermined threshold value of the voiced sound signal taken in by the voiced sound taking means Means for inverting the polarity of the peak value detected by the peak detecting means, adding the result to the extreme value detected by the extreme value detecting means, and performing a clamp process; Pitch period extracting means for extracting a pitch period based on a voice signal.

According to a second aspect of the present invention, there is provided a pitch period extracting apparatus comprising:
Pitch period extracting means, detecting means for detecting the voiced sound signal clamped in the clamp processing means,
Differentiation processing means for differentiating the waveform detected by the detection means. The pitch period extracting device according to claim 3 further includes a blocking unit that blocks the passage of a voiced sound signal having a period shorter than the pitch period of a general person's voice before the peak detecting unit.

According to a fourth aspect of the present invention, there is provided a pitch period extracting apparatus comprising:
A preceding stage of the peak detecting means further comprises AGC processing means for performing AGC processing on the voiced sound signal when the rising amplitude and the falling amplitude of the voiced sound signal captured by the voiced sound capturing means are small. . The pitch period extracting device according to claim 5 includes both a cutoff unit and an AGC processing unit at a stage preceding the peak detection unit.

A pitch period extracting apparatus according to claim 6 is
Peak detecting means for detecting a peak existing in one area outside a predetermined width of one area from a predetermined threshold, and extreme value detecting means detecting a peak in the other area further from the predetermined threshold. It detects an extreme value existing in the other area outside the width. The method for extracting a period according to claim 7, wherein the voiced sound signal is captured as time-series data, and a peak present in one of a predetermined threshold value of the voiced sound signal captured in the voiced sound capturing step. An extreme value detecting step of detecting an extreme value existing in the other region of the predetermined threshold value of the voiced sound signal captured by the voiced sound capturing step,
A clamp processing step in which the polarity of the peak value detected in the peak detection step is inverted and added to the extreme value detected in the extreme value detection step to perform the clamp processing; Extracting a pitch period based on the pitch period.

In the cycle extracting method according to the present invention, the pitch cycle extracting step includes detecting a voiced sound signal clamped in the clamping step, and differentiating the waveform detected in the detecting step. Processing stage. In the cycle extracting method according to the ninth aspect, prior to performing the peak detecting step, a blocking step of blocking the passage of a voiced sound signal having a cycle shorter than the pitch cycle of the voice of a general person is further performed.

According to a tenth aspect of the present invention, in the method of the present invention, when the amplitude of the rising edge of the voiced sound signal captured in the voiced sound capturing step is small and the amplitude is small prior to the execution of the peak detecting step, the AGC is added to the voiced sound signal. AG to process
Perform the C processing step further. In the cycle extracting method according to the eleventh aspect, both the cutoff step and the AGC processing step are performed before the peak detection step is performed.

According to a twelfth aspect of the present invention, in the peak extracting step, the peak detecting step detects a peak present in one area outside a predetermined width of the other area from a predetermined threshold value. The step is to detect an extreme value existing in the other area outside the other area by a predetermined width from the predetermined threshold value.

[0013]

FIG. 1 is a block diagram of a pitch period extracting apparatus according to the present invention. A voiced sound converted into an electric signal by a microphone 11 is taken into a microcomputer 13 via a line amplifier 12. The microcomputer 13 has a data bus 131 as a center,
It comprises a CPU 132, a memory 133, an A / D converter 134 and an output interface 135.

The extracted pitch period is output to an output device 1 such as a printer via an output interface 135.
4 is output. FIG. 2 is a flowchart of a first pitch cycle extraction routine executed by the microcomputer 131.
The voiced sound signal S collected in step 1 is read. It is assumed that the voiced sound signal S is biased by the line amplifier 12 so as to be read as a signal that oscillates positively and negatively around 0 volt.

In step 22, peak detection processing is executed, and in step 23, extreme value detection / clamping processing is executed. Further, after performing the pitch cycle detection processing in step 24, the pitch cycle is output in step 25, and this routine ends. FIG. 3 is a detailed flowchart of the peak detection process executed in step 22. In step 221, it is determined whether or not a peak is detected in the voiced sound signal S in the positive region.

If the determination in step 221 is affirmative, that is, if a peak is detected, the process proceeds to step 222, where the peak value in the positive direction is stored as P, and this routine ends. If a negative determination is made in step 221, that is, if no peak is detected, this routine is directly terminated. FIG. 4 is a detailed flowchart of the extremum detection / clamping processing routine executed in step 23. In step 231, it is determined whether or not an extremum is detected in the voiced sound signal S in a negative region.

If an affirmative determination is made in step 231, that is, if an extreme value is detected, the process proceeds to step 232, in which a value obtained by subtracting the peak value P from the voiced sound signal S is set as the processing value X, and this routine ends. I do. If a negative determination is made in step 231, that is, if no extreme value is detected, the process proceeds to step 233, where the voiced sound signal S is set as the processing value X, and this routine ends.

FIG. 5 is a detailed flowchart of the pitch period detection process executed in step 24. In step 241, the detection output Y 'is updated by the following equation. Y ′ = E · exp (−Δt / τ) where Δt is a sampling time and τ is a predetermined time constant. E will be described later.

In step 242, it is determined whether or not the absolute value of the processing value X is larger than the absolute value of the detection output Y '. If a negative determination is made in step 242, that is, if no peak is detected in the processing value X, the process proceeds to step 243, the detection output Y 'is set to E, and the process proceeds to step 246.

If an affirmative determination is made in step 242, the process proceeds to step 244, in which it is determined whether or not an extreme value has been detected in the negative region of the voiced sound signal S. If an affirmative determination is made in step 244, the processing value X is set to E in step 245,
Proceed to step 246. If a negative determination is made in step 244, the process proceeds to step 243.

In step 246, the value stored as E is set as the detection output Y, and in step 247, the detection output Y is differentiated by the following equation. Q = (Y−YB) / Δt In step 248, the detection output Y is prepared for the next calculation.
Is replaced with the previous detection output YB, and this processing is terminated.

That is, the pitch period can be detected by measuring the period of the differential value Q whose absolute value is larger than a predetermined value. FIG. 6 is a diagram for explaining the reason why filtering processing is performed on a voiced sound signal having a cycle shorter than the pitch pitch of a general person performed before step 22, as shown in (a) and (b). In order to reliably detect the peak a in the positive region or the peak b in the negative region, it is necessary to reliably remove the extreme value c in the middle.

In the present invention, the minimum value of the peak a in the positive region and the minimum value of the negative region after the peak a in the positive region, that is, c in (b) is detected, and the polarity of the peak a is inverted to obtain the negative region Is added to the minimum value of According to this processing, even when the positive and negative amplitudes are the same as shown in FIG.

FIGS. 7 to 10 are waveform diagrams of voiced sound processing by the first pitch period extraction routine. FIG. 7 is a waveform diagram of voiced sound collected by the microphone 11. FIG.
Is a waveform diagram after the clamp process, in which the peak value P is detected at times t ₁ , t ₃ and t ₅ , and the polarity is inverted to the extreme values detected at times t ₂ , t ₄ and t _6. Value P
Are added to perform a filtering process.

FIG. 9 is a waveform diagram after the detection process, and FIG. 10 is a waveform diagram after the differentiation process. The pitch period can be measured by measuring the pulse period after the differentiation process. FIG. 6 also shows the first pitch period extraction routine.
As shown in (b), when the absolute value of the minimum value in the negative region is small, the extreme value cannot be sufficiently enhanced by the filtering process.

In order to solve this problem, it is effective to filter voiced sounds to remove short-period voiced sounds. FIG. 11 is a flowchart of the second pitch period extraction routine. A filtering process 111 is inserted between Steps 21 and 22 of the first pitch period extraction routine.

In this filtering process, it is possible to use a well-known digital filter having a low-pass characteristic. It is preferable that the cutoff frequency is set so as to cut off signals of 300 to 400 Hz or more, which is the pitch frequency of ordinary people. FIG. 12 is an explanatory diagram of the effect of the filtering process. FIG.
(B) shows the processed waveform.

According to the first and second pitch cycle extraction routines, it is possible to extract the pitch cycle if the voiced sound has an amplitude greater than or equal to a predetermined value, but if the amplitude is small, the extracted pitch cycle is used. Accuracy may be degraded. FIG. 13 is a flowchart of a third pitch cycle extraction routine to which an AGC process for solving this problem is added. An AGC process 131 is inserted between steps 21 and 22 of the first pitch cycle extraction routine. You.

FIG. 14 is a detailed flowchart of the AGC process. In step 131A, it is determined whether or not the voiced sound signal S has a maximum value. If the determination is affirmative, the maximum value is stored as _Smax in step 131B. . Further, in step 131C, a parameter h representing the time after the detection of the maximum value is set, and the sampling number i representing the time at which the maximum value is detected is set as i _max (h) in step 13C
Go to 1F.

If a negative determination is made in step 131A, the flow advances to step 131D to determine whether or not the voiced sound signal S has a minimum value. If an affirmative determination is made in step 121D, the minimum value is stored as S _min in step 131E, and the flow proceeds to step 131F. Step 131F
Then, the amplitude ΔS of the voiced sound signal S is calculated by the following equation.

ΔS = | S _max −S _min | In step 131G, the moving average value ΔS _av of the amplitude is calculated as a function of the amplitude ΔS by the following equation, and the routine proceeds to step 131H. ΔS _av = {(n−1) · ΔS _av + ΔS} / n If a negative determination is made in step 131D, the process directly proceeds to step 131H.

In step 131H, the AGC gain AG (h) is determined as a function of the moving average value ΔS _av of the amplitude in the following equation. AG (h) = AG (ΔS _av ) Specifically, the function of the following equation can be used. AG (h) = A · exp (−ΔS _av / V) + B where A, V and B are predetermined constants.

FIG. 15 is a graph showing an example of the function of the AGC gain. The horizontal axis represents the moving average value ΔS _av of the amplitude, and the vertical axis represents the AGC gain. As shown in this graph, the reason why the AGC gain sharply decreases with the increase in the moving average value ΔS _av of the amplitude is to compensate for the saturation characteristic of voiced sound to make the amplitude uniform. In step 131I, the parameter h representing the time after the detection of the local maximum value is incremented, and in step 131J, the sampling number (i-Δi) that is retroactive from the sampling number i representing the current time by the AGC gain calculation time Δi is i _max (0 ) Is determined.

When an affirmative determination is made in step 131J, the routine proceeds to step 131K, where the compensation gain G is set to AG (0). Then, in step 131L, the AGC gain AG
(J) and the sampling number i _max (j) representing the time at which the maximum value is detected are shifted forward by the following equation,
Proceed to step 131M.

AG (j-1) = AG (j) i _max (j-1) = i _max (j) where j = 1, 2,... H Note that if a negative determination is made in step 131J, step 131M is directly performed. Proceed to. In step 131M, AGC processing is executed based on the following equation, and this routine ends.

S (i) = GS (i-Δi) The filtering process used in the second pitch period extraction routine can be added before the AGC process in step 131. FIG. 16 is a graph showing a pitch cycle extraction result by the third pitch cycle extraction routine, in which a pitch cycle is extracted at about 150 Hz.

However, from 2 to 2.5 seconds, the pitch period is distributed over a wide frequency band even though there is no voiced sound. This has been found to be the effect of so-called background noise, and when the voiced sound signal is within a predetermined width centered on 0 volt, the voiced sound signal is forcibly reduced to 0.
A process for replacing bolts (hereinafter referred to as dead zone processing) has been added.

FIG. 17 is a flow chart of the fourth pitch cycle extraction routine, wherein step 17 is executed between step 21 and step 131 of the third pitch cycle extraction routine.
A dead zone process is inserted as 1. FIG. 18 is a detailed flowchart of the dead zone processing. In step 171A, it is determined whether or not the voiced sound signal S (i) is equal to or smaller than a predetermined value DB.

When an affirmative decision is made in step 171A, the operation proceeds to step 171B, where the voiced sound signal S (i) is generated.
This process is set to "0" and the process ends. Step 171
If a negative determination is made in A, the process proceeds to step 171C, and it is determined whether or not the voiced sound signal S (i) is equal to or greater than a predetermined value -DB. When an affirmative determination is made in step 171C, the process proceeds to step 171D, and the voiced sound signal S (i)
Is set to "0", and this processing ends.

If a negative determination is made in step 171C, this process ends directly. FIG. 19 is a graph showing a pitch cycle extraction result by the fourth pitch cycle extraction routine, in which a pitch cycle is extracted at about 150 Hz. Further, in a portion where no voiced sound of 2 to 2.5 seconds exists, no pitch period is detected.

[0041]

According to the pitch period extracting apparatus according to the first aspect and the pitch period extracting method according to the seventh aspect, the polarity of the peak value in one area is inverted and added to the extreme value in the other area. By performing the clamping process described above, it is possible to detect the pitch period in the time domain almost in real time.

According to the pitch period extracting apparatus and the pitch period extracting method of the present invention, the pitch period is reliably detected by detecting and differentiating the voiced sound after the clamping process and extracting the pitch period. Can be detected. According to the pitch cycle extracting device according to the third aspect and the pitch cycle extracting method according to the ninth aspect, the clamp process is performed after the cutoff process for cutting off a voiced sound having a cycle equal to or longer than the pitch cycle of a general person. It is possible to reliably detect an extreme value and to improve the detection accuracy of the pitch period.

According to the pitch period extracting apparatus according to the fourth aspect and the pitch period extracting method according to the tenth aspect, when the amplitude of a voiced sound is small, the amplitude is amplified by AGC processing to thereby improve the accuracy of detecting the pitch period. It is possible to improve. According to the pitch cycle extracting device according to claim 5 and the pitch cycle extracting method according to claim 11,
By performing both the low-pass filter processing and the AGC processing, it is possible to further improve the pitch cycle detection accuracy.

According to the pitch period extracting apparatus of the sixth aspect and the pitch period extracting method of the twelfth aspect, voiced sounds having an amplitude equal to or less than a predetermined dead band width are removed to improve the detection accuracy of the pitch period. It becomes possible.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a pitch period extracting device.

FIG. 2 is a flowchart of a first pitch cycle extraction routine.

FIG. 3 is a detailed flowchart of a peak detection process.

FIG. 4 is a detailed flowchart of an extreme value detection / clamping process.

FIG. 5 is a detailed flowchart of a pitch cycle detection process.

FIG. 6 is an explanatory diagram of a reason for performing a filtering process.

FIG. 7 is a waveform diagram of a voiced sound.

FIG. 8 is a waveform diagram after a clamp process.

FIG. 9 is a waveform diagram after a detection process.

FIG. 10 is a waveform diagram after a differentiation process.

FIG. 11 is a flowchart of a second pitch cycle extraction routine.

FIG. 12 is an explanatory diagram of an effect of a filtering process.

FIG. 13 is a flowchart of a third pitch cycle extraction routine.

FIG. 14 is a detailed flowchart of an AGC process.

FIG. 15 is a graph showing an example of a function of an AGC gain.

FIG. 16 is a graph showing an extraction result obtained by a third pitch period extraction routine.

FIG. 17 is a flowchart of a fourth pitch cycle extraction routine.

FIG. 18 is a detailed flowchart of a dead zone process.

FIG. 19 is a graph showing an extraction result obtained by a fourth pitch period extraction routine.

[Description of Signs] 11 Microphone 12 Line amplifier 13 Microcomputer 131 Data bus 132 CPU 133 Memory 134 A / D converter 135 Output interface 14 Printer

Continued on the front page (72) Inventor Kazuo Takayama 1-2-2, Goshodori, Hyogo-ku, Kobe, Hyogo Prefecture Inside Fujitsu Ten Co., Ltd. (72) Inventor Atsushi Takashima 1-2-1, Goshodori, Hyogo-ku, Kobe-shi, Hyogo No. 28 Inside Fujitsu Ten Limited (72) Inventor Toshitaka Yamato 1-2-2 Goshodori, Hyogo-ku, Kobe City, Hyogo Prefecture Inside Fujitsu Ten Limited (56) References JP-A-2-114300 (JP, A) JP-A-60-191298 (JP, A) JP-A-55-55398 (JP, A) JP-A-1-101600 (JP, A) JP-A-61-26098 (JP, A) JP-A 64-82098 ( JP, A) Japanese Utility Model Showa 55-11523 (JP, U) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/04

Claims (12)

(57) [Claims] 1. A voiced sound capturing means for capturing a voiced sound signal as time-series data, and a peak for detecting a peak present in one of predetermined threshold regions of the voiced sound signal captured by the voiced sound capturing means. Detecting means, an extreme value detecting means for detecting an extreme value existing in the other area of the predetermined threshold value of the voiced sound signal fetched by the voiced sound fetching means, and a peak value detected by the peak detecting means And a clamp processing means for performing a clamp process by adding the value to the extreme value detected by the extreme value detection means, and extracting a pitch period based on the voiced sound signal clamped by the clamp processing means. And a pitch period extracting means for extracting a pitch period of a voiced sound signal. 2. The detecting device according to claim 1, wherein the pitch period extracting unit includes: a detecting unit configured to detect the voiced sound signal clamped by the clamping unit; and a differential processing unit configured to perform a differential process on a waveform detected by the detecting unit. The apparatus for extracting a pitch period of a voiced sound signal according to claim 1. 3. The pitch period of the voiced sound signal according to claim 1, further comprising a blocking unit that blocks passage of a voiced sound signal having a period shorter than a pitch period of a general person's voice, before the peak detecting unit. Extraction device. 4. An AGC process is performed on the voiced sound signal before the peak detecting means when the rising amplitude and the falling amplitude of the voiced sound signal captured by the voiced sound capturing means are small. 2. The pitch period extracting apparatus for a voiced sound signal according to claim 1, further comprising AGC processing means. 5. The voiced signal pitch period extracting apparatus according to claim 1, further comprising: both the cutoff means and the AGC processing means at a stage preceding the peak detection means. 6. The apparatus according to claim 1, wherein the peak detecting means detects a peak present in one area outside a predetermined width of the one area from a predetermined threshold value, and the extreme value detecting means determines a predetermined threshold value. 2. The pitch period extracting device for a voiced sound signal according to claim 1, wherein an extreme value existing in the other area outside the other area by a predetermined width from the threshold value is detected. 7. A voiced sound capturing step for capturing a voiced sound signal as time-series data, and a peak for detecting a peak existing in one of predetermined threshold values of the voiced sound signal captured by the voiced sound capturing step. A detecting step, an extreme value detecting step of detecting an extreme value existing in the other region of the predetermined threshold value of the voiced sound signal captured by the voiced sound capturing step, and a peak value detected in the peak detecting step Inverting the polarity and adding the extreme value detected in the extreme value detecting step to perform a clamp process; and extracting a pitch period based on the voiced sound signal clamped in the clamp process stage. Extracting a pitch period of a voiced sound signal. 8. The pitch period extracting step includes: a detecting step of detecting a voiced sound signal clamped in the clamping step; and a differential processing step of differentiating a waveform detected by the detecting step. The method for extracting a pitch period of a voiced sound signal according to claim 7. 9. A blocking step of blocking passage of a voiced sound signal having a period shorter than a pitch period of a general person's voice before performing the peak detecting step.
2. A method for extracting a pitch period of a voiced sound signal according to item 1. 10. The method according to claim 1, wherein prior to performing the peak detecting step,
8. The voiced sound signal according to claim 7, further comprising performing an AGC processing step of performing an AGC process on the voiced sound signal when the rising amplitude of the voiced sound signal captured in the voiced sound capturing step is small and the amplitude is small. Pitch period extraction method. 11. Prior to performing the peak detection step,
The method according to claim 7, wherein both the blocking step and the AGC processing step are performed. 12. The peak detecting step detects a peak present in one area outside a predetermined width of one area further from a predetermined threshold value, and the extreme value detecting step includes a predetermined threshold value. 8. The method for extracting a pitch period of a voiced sound signal according to claim 7, further comprising detecting an extreme value existing in the other area outside the other area by a predetermined width from the threshold value.