JP2532618B2 - Pitch extractor - Google Patents

Abstract

PURPOSE:To decide a fact that a pitch exists in one of a positive or negative signal and to extract the pitch in a voice period by driving a signal into a positive signal part and a negative signal part by a positive/negative deciding part, binarizing the positive signal or the negative signal in accordance with an energy level of the signal is large or small by a binarizing part, and deriving an auto-correlation of each signal which is binarized. CONSTITUTION:A signal from a voice input terminal 1 passes through an LPF 2 and inputted to a positive/negative deciding part 3, a waveform peak level of every unit time of the signal is detected, and thereafter, the signal is divided into a positive signal part and a negative signal part, and the positive signal part is inputted to a binarizing part 4. In this binarizing part 4, a prescribed level of the waveform peak level is set to a clip level, and binarized by determining '1'.'0'. Subsequently, a correlation of each signal which is binarized is derived by a correlation arithmetic part 5, and inputted to a pitch existence deciding part 6. Next, when the correlation exists in one of the positive signal part and the negative signal part, it is decided that a pitch exists by the deciding part, the pitch is extracted in a voice section, and an erroneous extraction of the pitch in other signal section than the voice section is decreased.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch extracting device.

2. Description of the Related Art In recent years, with the development of voice recognition technology, voice recognition devices are being put to practical use in various fields. In order to put it into practical use, it is necessary to solve various problems in actually using the recognition device. One of the problems in this use is that when a signal other than voice (for example, various environmental noises or breath sounds that are likely to be input before and after voice) is input to the recognition device with high power, However, there is a point that an incorrect recognition result is output because only the voice section cannot be extracted. In order to solve the above problems, it is necessary to extract a voice section using information other than the presence / absence of power, and it is possible to use the pitch presence / absence information for section extraction by using the pitch extraction device for a voice recognition device. Attempted,
There is a problem that the pitch may not be extracted in the voice part when voice and noise are input at the same time.

In order to solve the above-mentioned problems, the conventional pitch extracting device digitizes the input signal into three values according to the level of the signal, and
The presence / absence of pitch was determined by the presence / absence of correlation in the digitized signal. By the above method, the pitch can be extracted by the voice unit even when voice and noise are simultaneously input.

Hereinafter, the conventional pitch extracting apparatus as described above will be described with reference to the drawings. FIG. 5 is a block diagram of a conventional pitch extracting device. 1 is a voice input terminal, 1
Reference numeral 0 is an LPF (Low Pass Filter) unit, 11 is a ternarization unit, 12 is a correlation calculation unit, 13 is a pitch presence / absence determination unit, and 7 is a pitch extraction result output terminal. The operation of the pitch extracting device configured as described above will be described below.

First, the signal input from the audio input terminal 1 is the LPF section 1
The signal input to 0 and passed through the LPF in the LPF unit 10 (for example, cutoff frequency 900 Hz) is input to the ternarization unit 11. The ternarization unit 11 finds the peak level of the absolute value of the signal for each unit time, sets the level of a fixed multiple (eg, 70%) of the peak level as the clip level, and for the positive signal, the clip level equal to or higher than 1, The following values are set to 0, the negative signal is set to −1 when the absolute value of the signal is equal to or higher than the clip level, and the value is 0 when the clipping level is 0 or less. As a method of searching the peak level, for example, when the unit time is set to 24 msec, the smaller one of the maximum levels of the absolute value of the signal at the beginning and the end of the unit section of 8 msec may be set as the peak level.

Next, the correlation calculator 12 obtains the autocorrelation of the ternary signal. For example, if the unit section length is 24 msec with a signal having a sampling frequency of 8000 Hz, the correlation coefficient of the 192nd order is calculated by calculating the correlation of the section total length.

Next, in the pitch presence / absence determining unit 13, if the ratio between the first peak level and the second peak level from the smallest order of the correlation coefficient is a certain value or more (for example, 30% or more), a corresponding unit It is considered that the section has a pitch, and if it is below, it is considered that there is no pitch, and the pitch extraction result is output from the output terminal 7.

However, in the configuration as described above, the pitch extraction performance is good when the S / N is relatively high, but when the S / N is low and the noise level is higher than the clip level, It becomes difficult to remove noise by ternarization, and pitch extraction is not performed correctly in the voice section. In addition, when the clip level is set to a high value in accordance with the noise level, the maximum crest value in one pitch and the waveform on the same sign side often become a binarized wave with a period for each pitch, but the waveform on the different sign side. However, since it becomes unstable with respect to the clip level, there is a problem in that the correlation between the voice sections becomes low and the pitch cannot be extracted.

In view of the problems of the related art as described above, the present invention provides a pitch capable of extracting a pitch in a voice section even when the S / N is low and the noise level is higher than a signal level other than the maximum peak value within one pitch. An object is to provide an extraction device.

Means for Solving the Problems The present invention according to claim 1 provides a positive / negative judging section for judging the positive / negative of an input signal, a binarizing section for binarizing a positive signal section or a negative signal section, and a correlation for obtaining a correlation between signals. It is composed of a computing unit and a pitch presence / absence determining unit for determining the presence / absence of a pitch.

The present invention according to claim 2 determines a positive / negative determination unit for determining positive / negative of an input signal, a binarization unit for binarizing a positive signal unit or a negative signal unit, a correlation calculation unit for obtaining a signal correlation, and a pitch candidate. The pitch candidate determination unit and the pitch presence / absence determination unit that determines the presence / absence of a pitch are configured.

With this configuration, the pitch extraction device of the present invention according to claim 1 divides the signal into a positive signal portion and a negative signal portion by the positive / negative determination portion,
The binarization unit binarizes the positive signal or the negative signal according to the magnitude of the energy level of the signal, and the correlation calculation unit obtains the autocorrelation of each binarized signal. If there is a correlation in either one of the signals, it is determined that there is a pitch.

In the pitch extraction device of the present invention according to claim 2, the positive / negative determination section divides the signal into a positive signal section and a negative signal section, and the binarization section outputs a positive signal.
Or, a negative signal is set to 2 depending on the magnitude of the energy level of the signal.
The correlation is calculated for each of the signals binarized by the correlation calculation unit, and if there is no correlation between the positive signal and the negative signal in the pitch candidate determination unit, there is no pitch, and if there is correlation in either one, the pitch is calculated. A candidate is determined, and if the pitch candidate is appropriate as a pitch of voiced speech, the pitch presence / absence determining unit determines that there is a pitch, and if it is inappropriate, determines that there is no pitch.

Embodiment Hereinafter, an embodiment of the pitch extracting apparatus of the present invention will be described with reference to the drawings.

Embodiment 1 FIG. 1 is a block diagram of a pitch extracting device in an embodiment of the present invention. 1 is audio input terminal, 2 is LPF (Low Pas
s Filter) section, 3 is a positive / negative determination section, 4 is a binarization section, 5 is a correlation calculation section, 6 is a pitch presence / absence determination section, and 7 is a pitch extraction result output terminal. Have been given a number.

The operation of the pitch extracting device configured as described above will be described below.

First, the signal input from the audio input terminal 1 is (3rd
(Step 21 in the figure) is input to the LPF unit 2, and the LPF unit 2 inputs the LPF.
The signal that has passed the cutoff frequency of 900 Hz (for example, step 22 in FIG. 3) is input to the positive / negative determination section 3.
In the positive / negative determination unit 3, after finding the waveform peak level of the signal for each unit time, the signal is divided into a positive signal unit and a negative signal unit,
First, the positive signal portion for each unit time is input to the binarizing portion 4 (FIG. 3, step 23). The peak level search method is
For example, if the unit time is 24 msec, the smaller of the maximum waveform levels of the first and last 8 msec of the unit section may be the peak level.

The binarizing unit 4 binarizes the signal by setting a level that is a certain multiple (for example, 70%) of the waveform peak level as a clip level, setting the clip level equal to or higher than 1 and setting the clip level equal to or lower than 0 (FIG. 3, step 24). ). Next, the correlation calculation unit 5 obtains the autocorrelation of the binarized signal (step 2 in FIG. 3).
Five). For example, if the unit interval length is 24 msec for a signal with a sampling frequency of 8000 Hz and the correlation of the entire interval length is calculated,
A second-order correlation coefficient is calculated.

Next, in the pitch presence / absence determining unit 6, a unit corresponding to a case where the ratio of the first peak value to the second peak value from the smallest order of the correlation coefficient is a certain value or more (for example, 30% or more). It is determined that there is a pitch in the section, and the result is output from the pitch extraction result output terminal 7 (FIG. 3, step 27).

When the ratio of the peak values is equal to or less than a certain value, the negative signal part within the same unit time as the above process is input to the binarizing part (third value).
FIG. 28, step 28), and the above-described processing after the binarization part is performed (FIG. 3, steps 29 and 30), and the pitch presence / absence judgment part 6 calculates the first peak and the second peak value of the correlation coefficient. When the ratio is equal to or more than a certain value, it is determined that there is pitch, and when the ratio is less than the certain value, it is determined that there is no pitch (FIG. 3, step 31), and the result is input to the pitch extraction result output terminal 7 (FIG. 3, step).
32).

As described above, according to the present embodiment, the positive / negative determination unit divides the signal into the positive signal unit and the negative signal unit, and the binarization unit binarizes the positive signal or the negative signal according to the magnitude of the energy level of the signal. , S / N is calculated by obtaining the autocorrelation of each binarized signal in the correlation calculation unit and determining that there is a pitch if the positive or negative signal is correlated in the pitch presence / absence determination unit. Even if the noise level is low and higher than the signal level other than the maximum peak value within one pitch, the pitch can be extracted in the voice section.

Second Embodiment FIG. 2 is a block diagram of a pitch extracting device according to a second embodiment of the present invention. 1 is audio input terminal, 2 is LPF (Low P
ass Filter) section, 3 is a positive / negative determination section, 4 is a binarization section, 5 is a correlation calculation section, 8 is a pitch candidate determination section, 9 is a pitch presence / absence determination section, and 7 is a pitch extraction result output terminal. The same items as the examples are given the same numbers.

The operation of the pitch extracting device configured as described above will be described below.

First, the signal input from the voice input terminal 1 is (4th
In the figure, step 41), it is input to the LPF unit 2, and the LPF unit 2 inputs the LPF.
A signal that has passed a cutoff frequency of 900 Hz (for example, step 42 in FIG. 4) is input to the positive / negative determination section 3. The positive / negative determination section 3 finds the waveform peak level of each signal per unit time, then divides the signal into a positive signal section and a negative signal section, and first inputs the positive signal section per unit time to the binarization section 4 ( Figure 4, step 43). The peak level searching method is the same as that in the first embodiment.

The binarizing unit 4 binarizes the signal by setting a level that is a certain multiple (for example, 70%) of the waveform peak level as a clip level, setting the clip level equal to or higher than 1 and the clip level equal to or lower than 0 (FIG. 4, step 44). ). Next, the correlation calculator 5 obtains the autocorrelation of the binarized signal (step 4 in FIG. 4).
Five). For example, if the unit interval length is 24 msec for a signal with a sampling frequency of 8000 Hz and the correlation of the entire interval length is calculated,
A second-order correlation coefficient is calculated.

Next, a unit corresponding to a case where the ratio of the first peak value to the second peak value from the smallest order of the correlation coefficient is a certain value or more (for example, 30% or more) in the pitch candidate determining unit 8. It is determined that there is a pitch in the section, and the result is input to the pitch presence / absence determining unit 9 with the reciprocal of the distance between peaks (unit: second) being the pitch (FIG. 4, steps 46 and 47).

When the ratio of the peak values is equal to or less than a certain value, the negative signal part within the same unit time as the above process is input to the binarizing part (fourth value).
(FIG. 48, step 48), and the above-described processing after the binarization part is performed (FIG. 4, steps 49, 50), and the pitch presence / absence determination part 6 determines the first peak and the second peak value of the correlation coefficient. If the ratio is equal to or greater than a certain value, it is determined that there is a pitch, and if it is less than a certain value, it is determined that there is no pitch (step 51 in FIG. 4).
In the figure, step 52), if there is no pitch, the result is output from the pitch extraction result output terminal 7.

Next, the pitch presence / absence determining unit 9 determines whether the pitch determined in the above process is appropriate as the pitch of voice according to the following conditions (1) and (2) (FIG. 4, step 54).

(1) The difference between the current pitch candidate and the pitch candidate determined by the signal in the previous unit section is 20 Hz or less.

 (2) Pitch candidates are 40 Hz or more and 400 Hz or less.

Only when both of the above two conditions are satisfied, it is determined that there is a pitch, and when neither of them is satisfied, it is determined that there is no pitch and the result is output from the pitch extraction result output terminal 7.
(FIG. 4, step 55).

As described above, according to the present embodiment, the positive / negative determination unit divides the signal into the positive signal unit and the negative signal unit, and the binarization unit outputs the positive signal.
Or, a negative signal is set to 2 depending on the magnitude of the energy level of the signal.
The correlation is calculated for each of the signals binarized by the correlation calculation unit, and if there is no correlation between the positive signal and the negative signal in the pitch candidate determination unit, there is no pitch, and if there is correlation in either one, the pitch is calculated. When the candidate is determined and the pitch presence / absence determination unit determines that the pitch candidate is appropriate as the pitch of voiced speech, there is a pitch, and if it is inappropriate, there is no pitch.
Even if the noise level is low and the noise level is higher than the signal level other than the maximum peak value within one pitch, it is possible to extract the pitch in the voice section and further reduce erroneous extraction of the pitch in the signal section other than the voice.

Advantageous Effects of the Invention The present invention divides a signal into a positive signal portion and a negative signal portion in the positive / negative determination unit, binarizes a positive signal or a negative signal in the binarization unit according to the magnitude of the energy level of the signal, and in the correlation calculation unit. The S / N is low and the noise level is 1 by obtaining the autocorrelation of each binarized signal and determining whether there is a pitch if the positive or negative signal is correlated in the pitch presence / absence determining unit. The present invention realizes an excellent extraction device that can extract a pitch in a voice section even when the signal level is higher than the maximum peak value in the pitch.

Further, according to the present invention, the positive / negative determination unit divides the signal into a positive signal unit and a negative signal unit, and the binarization unit binarizes the positive signal or the negative signal according to the magnitude of the energy level of the signal, and the correlation calculation unit performs the binary conversion. The correlation of each binarized signal is obtained, and if there is no correlation with either the positive signal or the negative signal in the pitch candidate determination unit, there is no pitch, and if there is a correlation with either one, the pitch candidate is determined and the presence / absence of pitch is determined. If the pitch candidate is appropriate as a pitch of voiced speech in the section, there is a pitch, and if it is inappropriate, there is no pitch, so that the S / N is low and the noise level is 1
To realize an excellent pitch extraction device capable of extracting a pitch in a voice section and reducing false extraction of a pitch in a signal section other than voice even if the signal level is higher than a signal level other than the maximum peak value in the pitch. Is.

[Brief description of drawings]

FIG. 1 is a block diagram of a pitch extracting device according to an embodiment of the present invention, FIG. 2 is a block diagram of a pitch extracting device according to a second embodiment of the present invention, and FIG. 3 is the same first embodiment. FIG. 4, FIG. 4 is a flowchart of the second embodiment, and FIG. 5 is a block diagram of a pitch extracting device in a conventional example. 1 ... Voice input terminal, 2 ... LPF section, 3 ... Positive / negative judgment section, 4 ... Binarization section, 5 ... Correlation calculation section, 6, 9 ... Pitch presence / absence judgment section, 7 ... Pitch extraction Result output terminal, 8 ...
... Pitch candidate determination unit

Claims (2)

(57) [Claims] 1. A positive / negative judging section for judging whether the input signal is positive or negative, and a positive value portion of the signal (hereinafter referred to as a positive signal portion) or a negative value portion of the signal (hereinafter referred to as a negative signal portion) is binary. It has a binarizing unit for converting the signal, a correlation calculating unit for obtaining the correlation of the signal, and a pitch presence / absence determining unit for determining the presence / absence of pitch, and the positive / negative determining unit divides the signal into a positive signal unit and a negative signal unit, and binarizes the signal. Section binarizes a positive signal or a negative signal according to the magnitude of the energy level of the signal, the correlation operation section obtains the correlation of each of the binarized signals, and the pitch presence / absence determination section determines whether the signal is a positive signal or a negative signal. A pitch extraction device characterized by determining that there is a pitch if there is a correlation on one side. 2. A positive / negative judging section for judging whether the input signal is positive or negative, a binarizing section for binarizing the positive signal section or the negative signal section, a correlation calculation section for obtaining a signal correlation, and a pitch candidate for determining pitch candidates. The determination unit and the pitch presence / absence determination unit for determining the presence / absence of pitch are provided, and the positive / negative determination unit divides the signal into a positive signal unit and a negative signal unit, and the binarization unit determines a positive signal or a negative signal of the energy level of the signal. Binarization is performed according to the magnitude, and the correlation of each of the binarized signals is obtained by the correlation calculation unit. If there is no correlation between the positive signal and the negative signal in the pitch candidate determination unit, there is no pitch. If there is a pitch candidate,
A pitch extraction device characterized in that the pitch presence / absence determining unit determines that the pitch candidate is suitable as a pitch of a voiced sound, and determines that there is no pitch if it is inappropriate.