KR19990003629A - Apparatus and method for detecting pitch interval using second-order linear prediction residual and nonlinear filter - Google Patents

Abstract

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for detecting pitch information of a voiced sound signal, which is important information for processing a speech signal. In particular, a second linear prediction coefficient is obtained to quickly detect pitch information, and the main frequency component of the original signal is obtained through an inverse filter. After the residual signal is removed, the buffer 10 buffers the input signal in a given unit using a nonlinear system to emphasize the pitch component, and zero to second order autocorrelation from the data buffered in the buffer 10. In the autocorrelation coefficient unit 20 for obtaining the coefficients, the linear prediction coefficient unit 30 for obtaining the second-order linear prediction coefficients from the autocorrelation coefficients obtained in the autocorrelation coefficient unit 20, and the linear prediction coefficient unit 30 An inverse filter 40 that obtains a residual signal with respect to the original input signal by inversely filtering the obtained linear prediction coefficients, and a non-linearity with respect to the residual signal obtained by the inverse filter 40 Pitch period by comparing the peak value obtained by the nonlinear filter 50 to be filtered, the peak value of the signal filtered by the nonlinear filter 50, and the peak detector 60 and the peak detector 60 that detect the position thereof with a threshold value. It relates to a pitch interval detection device and method using a second-order linear prediction residual and a nonlinear filter composed of a pitch period determination unit 70 for determining the.

Description

Apparatus and method for detecting pitch interval using second-order linear prediction residual and nonlinear filter

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for detecting pitch information of a voiced sound signal, which is important information for processing a speech signal. In particular, a second linear prediction coefficient is obtained to quickly detect pitch information, and the main frequency component of the original signal is obtained through an inverse filter. The present invention relates to a pitch interval detection apparatus and method using a second-order linear prediction residual and a nonlinear filter, after obtaining a residual signal obtained by removing the residual signal, and then using a nonlinear system to emphasize the pitch component.

Pitch information of the voiced sound signal is important information for processing the voice signal.

Therefore, there have been many studies on how to find the pitch of voiced sounds.

There are methods using linear predictive residuals, methods using auto-correlation coefficients, and methods to obtain from 켑 strum coefficients.

In the method using the linear prediction residual, first, all poles are modeled using a linear filter having only a pole, and the model parameters are obtained from the input data.

This linear model models the characteristics of voice signals, especially voiced sound signals, with several different resonance points, and models these resonance points (formers in voiced sounds) as poles.

Thus, removing these poles removes the formant characteristics of the speech, resulting in a flat frequency characteristic.

In this way, in order to eliminate the pole, the inverse filter of the linear filter obtained above is applied to the input audio signal to obtain a residual.

In the residual signal obtained as described above, only continuous pulses of the pitch component remain in the voiced sound section, and only white noise remains in the unvoiced sound section.

Therefore, the positions of the peak values corresponding to the pitch components in this residual signal are found, and the difference between the peak positions is taken as the pitch period.

In general, as shown in FIG. 1, a pitch detection apparatus using a linear prediction filter includes a buffer 1 for storing an input signal and an autocorrelation coefficient of order 0 to 20 from data stored in the buffer 1. A linear prediction coefficient unit 3 for obtaining a linear prediction coefficient of order 20 from the autocorrelation coefficients obtained by the autocorrelation coefficient unit 2 and the linear prediction coefficient unit 3 obtained by the linear prediction coefficient unit 3 An inverse filter (4) for obtaining a residual signal with respect to the original input signal from prediction coefficients, a peak detector (5) for obtaining a peak value and its position from the residual signal obtained from the inverse filter (4), and the peak detector (5) It consists of an interval calculating section 6 for determining the pitch period by comparing the peak value obtained by the threshold value.

The conventional pitch detection device configured as described above operates as follows.

First, the linear prediction coefficient is mainly obtained by autocorrelation coefficient in consideration of the predicted filter stability.

Therefore, the input signal is first stored in the buffer 1 by one block unit (usually 20 to 30 msec), and the autocorrelation coefficient unit 2 obtains 0 to 20th order from the data of the buffer 1.

The linear prediction coefficient unit 3 obtains the 20th order linear prediction coefficient from the autocorrelation coefficients obtained as described above by the autocorrelation coefficient method.

The inverse filter 4 performs inverse filtering from the linear prediction coefficients obtained by the linear prediction coefficient unit 3 and applies the original input signal to obtain a residual signal.

With only the magnitude of the signal from the residual signal obtained by the inverse filter 4, the peak detector 5 finds the peak value and its position.

If the magnitude of the peak value obtained by the peak detector 5 exceeds a certain threshold, it is regarded as the pitch of the voiced sound, and the section calculator 6 determines the distance between the peak positions as the pitch period.

On the other hand, if it is smaller than the threshold, it is an unvoiced or silent section.

It is difficult to find the pitch in the above-described way because the pitch component is larger than the other components and can be easily distinguished, but the pitch component is small at the beginning and end of the voiced sound and the transition period.

In addition, the calculation amount is large, and the burden on the processor becomes too large to simply use only the pitch information.

That is, in the case of obtaining a pitch in units of 30 msec (240 samples) for an 8 kHz sampled signal, the multiplication calculation is as follows.

First, in order to find the autocorrelation coefficients of order 0 to 20, it is necessary to multiply approximately 21 x 240 = 5,040 times, and perform the inverse filter about 20 x log (20) ≒ 60 times to find the linear prediction coefficients of the 20th order. To do this we need to multiply 240 x 20 = 4,800 times.

That is, about 9,900 calculations are required in total, and a processor resource of 1 mips (hereinafter referred to as MIPS (MiPS)) is consumed.

On the other hand, the method using autocorrelation coefficient is a method of detecting the pitch by applying the property that the autocorrelation coefficient of the input signal has a peak value for every pitch period of the signal, which is larger than the method using the linear prediction residual. .

In other words, assuming that the pitch range is 70Hz to 300Hz when the autocorrelation coefficient is obtained, the autocorrelation coefficient of order 26 to 115 should be obtained.

Therefore, the number of multiplications used to find the autocorrelation coefficients is only 90 × 240 = 21,600, which requires 2.1 MIPS of resources.

The method using the Cepstrum coefficient utilizes a characteristic in which the Cepstrum appears separately from a source component on a signal generation model and a filter (a vocal model for voice signals).

That is, the components for the negative vocal filter are gathered at the low order of the Cepstrum, and the components corresponding to the source take advantage of the characteristics formed at the high order.

Thus, the pitch is found with peak values at higher orders (26 orders of magnitude or more when only pitch is considered up to 300 Hz).

In this case, when the fast Fourier transform (FFT) is used, it is necessary to multiply 256 x log 2 (256) = 2,048 times when using the 256th order, and to multiply 2,048 times in the inverse fast Fourier transform (FFT).

In addition, when determining the magnitude of the spectrum, 128 x 2 multiplications and 128 logarithmic operations are required.

In other words, it takes 4,350 multiplications and 128 log operations.

In general, the log operation is a multiple of several times the multiplication, and therefore, the cepstrum method requires about 1 mips (MIPS) of resources.

The present invention has been devised to solve the above-mentioned problems of the prior art, and obtains a linear linear prediction coefficient for quickly detecting the pitch information, and removes the residual frequency components of the original signal through the inverse filter. It is an object of the present invention to provide an apparatus and method for detecting a pitch interval using a second-order linear prediction residual and a nonlinear filter that emphasizes pitch components using a nonlinear system.

1 is a block diagram of a pitch detection apparatus using a conventional linear prediction residual;

2 is a block diagram of a pitch detection apparatus using a second-order linear prediction residual and a nonlinear filter according to the present invention;

3 is an exemplary diagram of a method for obtaining autocorrelation coefficients from buffer data;

4 is an exemplary diagram of a voice generation model;

5 is an exemplary waveform diagram of a beginning portion of voiced sound of a voice signal;

6 is an exemplary waveform diagram of a 20th order linear prediction residual for the speech signal of FIG. 5;

7 is an exemplary waveform diagram of a second-order linear prediction residual for the speech signal of FIG. 5;

8 is a waveform illustration of a nonlinear filter output for a second order linear prediction residual.

* Explanation of symbols for main parts of the drawings

10 buffer 20 autocorrelation coefficient unit

30: linear prediction coefficient unit 40: inverse filter

50 nonlinear filter 60 peak detection unit

70: pitch period determination unit 100: saint filter

200: continuous pulse 300: white noise

The present invention for achieving the object as described above, the autocorrelation coefficient to obtain a zero to second order autocorrelation coefficient from the buffer 10 for buffering the input signal in a given unit, and the data buffered in the buffer 10 A linear prediction coefficient unit 30 for obtaining a second linear prediction coefficient from the autocorrelation coefficients obtained by the autocorrelation coefficient unit 20 and an inverse of the linear prediction coefficients obtained by the linear prediction coefficient unit 30 An inverse filter 40 for filtering out a residual signal with respect to the original input signal, a nonlinear filter 50 for performing nonlinear filtering on the residual signal obtained in the inverse filter 40, and a filtering of the signal filtered by the nonlinear filter 50 And a pitch period determination unit 70 for detecting a peak value and its position, and a pitch period determination unit 70 for determining a pitch period by comparing the peak value obtained by the peak detection unit 60 with a threshold value. The.

In addition, the present invention for achieving the object as described above, the step of buffering the input signal until the data as much as the analysis interval, the samples obtained by moving the buffered data by 0 to 2 samples, and Obtaining the autocorrelation coefficient by multiplying and accumulating samples of each other, and obtaining the linear prediction coefficients {a1, a2} from the autocorrelation coefficients {r0, r1, r2} obtained using the autocorrelation method. A step of constructing an inverse filter with the linear prediction coefficients {a1, a2} obtained in the step, and obtaining a residual signal by passing the signal originally input to the buffer to the filter, using the nonlinear filter Filtering again, obtaining peak values and positions from the output of the nonlinear filter, and using the peak values and peak positions obtained in the step. For example, the method may include determining an average of each pitch section or determining a pitch based on a median average.

Pitch information of the voiced sound signal is important information for processing the voice signal.

In order to quickly extract this pitch information, in the present invention, the linear prediction coefficients of the second order are obtained, and the residual signal obtained by removing the main frequency components of the original signal is first obtained through the inverse filter.

The residual signal includes a pitch component, but also includes other formant information, so a nonlinear system is used to emphasize the pitch component.

First, the input signal is buffered until data for the analysis section is collected in the buffer 10.

This analysis interval is usually set to a length that can contain two or three pitch pulses, ranging from 20 msec to 30 msec, from 160 to 240 samples when the input signal is 8 kHz sampled.

As shown in FIG. 3, the autocorrelation coefficient unit 20 obtains samples obtained by moving data in the buffer 10 by 0 to 2 samples and accumulates the original samples by multiplying each other.

From the autocorrelation coefficients {r0, r1, r2} thus obtained, the linear prediction coefficient unit 30 obtains the linear prediction coefficients {a1, a2} using an autocorrelation method.

The inverse filter 40 has the linear prediction coefficients {a1, a2} obtained by the linear prediction coefficient unit 30, and configures the inverse filter first as follows.

[Equation 1]

Here, a0 is 1, x (n) is an audio signal which is an input of an inverse filter, and y (n) represents a residual signal which is an output of an inverse filter.

When the inverse filter is configured, the residual signal is obtained by passing the signal input to the original buffer through the filter.

The residual signal is again filtered out by the nonlinear filter 50.

The nonlinear filter 50 has a structure as follows.

[Equation 2]

Where z (n) is the output of the nonlinear filter and y (n) is the input of the nonlinear filter, i.e. the residual signal.

In the peak detector 60 that finds the peak value, the peak value and its position are obtained from the output of the nonlinear filter 50.

The peak value and the peak position are further determined by the pitch period determination unit 70.

Since there are two or three pitch pulses in one impeller, each pitch section is averaged or a median average when determining pitch.

Looking at the required calculation amount of the pitch detection device operating as described above is as follows.

First, 3 x 240 = 720 times to get the autocorrelation coefficient, several times to get the linear prediction coefficient, and 2 x 240 = 480 times to the inverse filtering, and 2 x 240 = 480 times to the nonlinear filtering. Multiplying is required, which requires approximately 1,700 multiplications in total, thus requiring only about 0.2 mips of process resources.

Here, the operation principle of the pitch interval detection apparatus using the second-order linear prediction residual and the nonlinear filter according to the object of the present invention will be described in detail.

In general, the generation of the voice signal is described as a source-filter model, as shown in FIG. 4, wherein the voice signal is generated by the vocal tract filter (the source signal from the gate). Because it comes through 100).

Here, the excitation signal is formed as a continuous pulse train 200 in the case of voiced sound, and has a form of white noise 300 in the case of unvoiced sound.

In addition, the duct filter 100 exhibits the characteristics of the phoneme, and is generally composed of formants having several resonance frequencies.

Therefore, in many cases, a linear prediction model is set for the speech signal and analyzed.

This is because the linear prediction model is mainly given as an all-pole model, and the frequency corresponding to this pole is similar to the formant of the duct filter 100.

In addition, the residual signal of the linear prediction model may be regarded as an excitation signal of speech generation.

Therefore, since a general speech generation model and a linear prediction model in speech analysis coincide with each other, they are widely used in speech signal related fields such as speech encoding, speech analysis, and speech recognition.

The electrode linear prediction model predicts the current sample x (n) with the past N samples {x (n-1), ..., x (nN)}, where the predicted current sample x '(n) is The linear combination of past N samples is obtained as

[Equation 3]

At this time, a prediction coefficient {ai, i = 1, .., N} indicating the linear coupling strength is obtained so that the error x (n) -x '(n) between the current sample and the predicted sample is small.

There are various ways to obtain prediction coefficients, and autocorrelation coefficients are widely used in consideration of predicted filter stability.

When the prediction coefficients are obtained by autocorrelation coefficients, N + 1 or more autocorrelation coefficients are required.

On the other hand, if the inverse filter is constructed from the linear prediction coefficients and applied to the original speech signal, the residual signal can be obtained. The residual signal is similar to the excitation signal of the speech generation model. It shows the form of noise.

Therefore, a method of detecting the pitch of voiced sounds by using the characteristics of the residual signal is disclosed.

That is, pitch detection of voiced sound using the residual signal is a method of removing a formant component in an audio signal to obtain an excitation signal and detecting a period of a pulse (pitch) from the excitation signal.

The inverse filter is constructed from the linear prediction coefficients {ai, i = 1, ..., N} as in the above equation (1).

Pitch detection using the residual signal of the linear prediction model generally consumes a lot of system resources because a linear prediction coefficient of about 20 orders has to be obtained.

Therefore, in the present invention, only the second-order linear prediction coefficient is obtained, and a residual obtained by removing only the main formant component that may be confused with the pitch component through the inverse filter is obtained, and the pitch period is detected therefrom.

At this time, in order to compensate the pitch component in the residual signal, a nonlinear filter such as Equation 2 is applied to the residual signal.

This nonlinear filter outputs a large value for the high energy portion of the input signal and a small value for the low energy portion, thus emphasizing the portion of the residual signal that corresponds to the high energy pulse and providing a low energy component. Decreases.

Therefore, it is advantageous to find the peak value when detecting the pitch of the voiced sound.

Hereinafter, an embodiment according to the present invention will be described with reference to FIGS. 5 to 8.

When pitch is detected in a speech signal, a lot of errors occur at the beginning and ending portions of the speech and the transition period.

Fig. 5 shows the beginning of the voice, and the peak value corresponding to the pitch is gradually increasing, and the shape is also changing a lot.

Next, FIG. 6 is a residual signal of the inverse filter by the linear prediction coefficient of the 20th order to show the performance of the method using the existing linear prediction residuals, but clearly shows the peak position, but the change in the signal level shows the original speech signal. Follow a lot

7 is a residual signal of the inverse filter by the second-order linear prediction coefficients used in the present invention, and there is not much difference from when the 20th-order linear prediction coefficients are used in the overall shape or performance.

Therefore, similar performance can be obtained by removing only the main formant component of the voice formant. Therefore, it is effective to reduce the consumption of system resources by obtaining the lower linear prediction coefficients.

8 is an output signal obtained by passing the residual signal of the inverse filter by the second order linear prediction coefficient to the nonlinear filter, and the pitch pulse component is clearly revealed.

Therefore, the performance degradation caused by using the second-order linear prediction coefficient may be sufficiently offset.

As described in detail above, the present invention is a method of detecting voice pitch using a method that consumes less resources of the system, and thus detects the detected voice quickly (more than five times faster than the method using the residual of the existing linear prediction model). It is suitable for the field available.

For example, in speech detection, speech analysis, speech recognition, speaker recognition, rhyme information extraction, etc., effective pitch detection is preferable rather than precise pitch information.

In the voice detection, it is possible to determine whether the input signal is a voiced sound using the presence or absence of the pitch, so that it is possible to determine whether the signal with high energy is voice or other dynamic noise.

In the voice analysis, the analysis section can be given a difference according to the pitch, so that the analysis can be performed more accurately.

On the other hand, in speech recognition, pitch information can be used to obtain a feature vector, and it can be used in continuous speech recognition using rhyme information.

Here, the pitch information is a component having the most information on the talker in the speech signal, so a method for quickly detecting the pitch is to add good information for speaker recognition.

Claims (4)

A buffer 10 for buffering an input signal in a given unit, an autocorrelation coefficient unit 20 for obtaining a 0 to 2nd order autocorrelation coefficient from the data buffered in the buffer 10, and the autocorrelation coefficient unit 20 The linear prediction coefficient unit 30 that obtains the second-order linear prediction coefficients from the autocorrelation coefficients obtained by, and the inverse that the linear prediction coefficients obtained by the linear prediction coefficient unit 30 are inversely filtered to obtain a residual signal with respect to the original input signal. A filter 40, a nonlinear filter 50 for nonlinear filtering the residual signal obtained by the inverse filter 40, and a peak detector 60 for detecting the peak value of the signal filtered by the nonlinear filter 50 and its position. And a pitch period determination unit 70 for determining a pitch period by comparing the peak value obtained by the peak detection unit 60 with a threshold value, and using a second-order linear prediction residual and a nonlinear filter. Detection device. Buffering the input signal until the data of the analysis period is collected, and obtaining autocorrelation coefficients by multiplying and accumulating the samples obtained by moving the buffered data by 0 to 2 samples, and the original samples. Obtaining the linear prediction coefficients {a1, a2} from the autocorrelation coefficients {r0, r1, r2} obtained in the step by using the autocorrelation method, and calculating the linear prediction coefficients {a1, a2} obtained in the above steps. And constructing an inverse filter, passing a signal originally input to the buffer to a filter to obtain a residual signal, re-filtering the residual signal obtained in the step with a nonlinear filter, and peak values from the output of the nonlinear filter. Using the step of finding the position and the peak value and the peak position obtained in the above step, the average of each pitch section or the average of the pitch The method of claim 1, characterized in that the step consisting of, the second-order linear prediction residual and the pitch interval detection method using a nonlinear filter. The method of claim 2, wherein the obtaining of the residual signal comprises: when a0 is 1, x (n) is an audio signal that is an input of an inverse filter, and y (n) represents a residual signal that is an output of an inverse filter. Pitch interval detection method using a second-order linear prediction residual and nonlinear filter, characterized in that using the same method. The method of claim 2, wherein the filtering in the nonlinear filter is performed when z (n) is an output of the nonlinear filter and y (n) is an input of the nonlinear filter, that is, a residual signal. Pitch interval detection method using a second-order linear prediction residual and nonlinear filter, characterized in that using the same method.