JP2001147700A - Method and device for sound signal postprocessing and recording medium with program recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reproduce a sound signal of high quality from an encoded bit sequence. SOLUTION: The postprocessing method of sound signal includes a stage where a sound signal is made to pass a linear filter 32 showing reverse characteristics of spectrum envelope, a stage where a peak position is detected from the signal passing the linear filter by a peak position detection part 33 and a waveform of one-pitch length is extracted by a signal waveform segmenting part 34, and a stage where peaks of mutual correlation are successively searched by a pitch reference position search part 35, and the signal which has passed the linear filter is divided into areas which correspond to the waveform of one-pitch length by a boundary area determination part 36. and a comb filter coefficient is calculated in a pitch correlation value calculation part 37, and the signal passing the linear filter in each area is made to pass a comb line filter 38 using this coefficient to obtain a sound signal output.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio signal post-processing method for improving the quality of a speech signal input for each frame by enhancing pitch components, and more particularly, to a filter for representing a spectral envelope characteristic of the audio signal. In predictive coding and decoding for driving a sound source vector to synthesize a voice, a post-processing method and apparatus effective for a voice signal applied to voice decoding for reproducing a high-quality voice signal from a coded bit sequence And a recording medium on which the program is recorded.

[0002]

2. Description of the Related Art In digital mobile communication, a high-efficiency voice coding method is used in order to efficiently use radio waves or to efficiently use a communication line or a storage medium for a voice or music storage service. ing. At present, as a method for encoding speech efficiently, original speech is called a frame or a subframe (hereinafter referred to as a frame).
The sound of one frame is divided into two information, that is, a linear filter characteristic representing an envelope characteristic of a frequency spectrum and a driving sound source signal for driving the filter. Has been proposed. In this method, as a method of encoding a drive excitation signal, a pitch period (basic frequency) of a voice is used.
There is known a method of separating and encoding a periodic component considered to correspond to the following and other components. Code-Excited Linear Prediction (CELP) is an example of an encoding method of the driving excitation information. For details of the above, see the document MR. Schroeder and B.
S. Atal, “Code-Excited Linear Prediction (CEL
P): High Quality Speech at Very Low Bit Rates ”, I
EEE Proc. ICASSP-85, pp. 937-940, 1985.

FIG. 1 shows an example of the configuration of the encoding unit 1. For the speech X input to the input terminal, the linear prediction analysis unit 2 calculates a linear prediction parameter representing a frequency spectrum envelope characteristic of the input speech. The obtained linear prediction parameters are quantized and coded in the linear prediction parameter coding unit 3, and the quantized parameters are sent to the synthesis filter 12 as synthesis filter coefficients k.

The details of the linear prediction analysis and examples of coding of the linear prediction parameters are described in, for example, “Digital Speech Processing” by Sadahiro Furui (Tokai University Press). Here, the linear prediction analysis unit 2, the linear prediction parameter coding unit 3, and the synthesis filter 12 may be replaced with non-linear ones. The driving sound source vector generation unit 5 generates a driving sound source vector candidate having a length of one frame, and sends the candidate to the synthesis filter 12. In the drive sound source vector generation unit 5, 1
A driving sound source vector candidate having a length corresponding to a frame is generated and sent to the synthesis filter 12. Driving excitation vector generation section 5 is generally composed of adaptive codebook 6 and fixed codebook 7 in many cases.
From the adaptive codebook 6, the immediately preceding past excitation vector stored in the buffer (the already-quantized immediately preceding one to several frames of excitation vectors) is cut out at a length corresponding to a certain period, and the cut out is performed. By repeating the generated vector until the frame length is reached, a time-series vector candidate corresponding to the periodic component of the voice is output. As the “certain period”, a period in which the distortion d in the distortion calculator 13 is reduced is selected. The selected period generally corresponds to the pitch period of the voice in many cases. The fixed codebook 7 outputs a time-series code vector candidate having a length of one frame corresponding to the non-periodic component of the voice. These candidates store a predetermined number of candidate vectors in accordance with the number of bits for encoding independently of the input speech. The time series vector candidates output from the adaptive codebook 6 and the fixed codebook 7 are multiplied by the weights created by the weight creation unit 10 in the multiplication units 8 and 9, respectively, added by the addition unit 11, and added to the driving sound source. This is a vector candidate c. In the configuration example of the driving excitation vector generation unit 5, the adaptive codebook 6 may not be used, and only the fixed codebook 7 may be used, and a signal having a small pitch periodicity such as a consonant part or background noise is encoded. In some cases, the adaptive codebook 6 is not used.

[0005] The synthesis filter 12 is a linear filter using the quantized value of the linear prediction parameter as a filter coefficient, and outputs a driving sound source vector candidate c as input and outputs a reproduced sound candidate y. In general, the order of the synthesis filter, that is, the order of the linear prediction analysis, is generally about 10 to 16 order. As described above, the synthesis filter 12 may be a non-linear filter. The distortion calculator 13 calculates a distortion d between the reproduced voice candidate y output from the synthesis filter and the input voice X. The calculation of this distortion is, for example,
In some cases, this is performed in consideration of the coefficients of the synthesis filter or the linear prediction coefficients that have not been quantized.

The codebook search control unit 14 selects a driving excitation code that minimizes the distortion d between each reproduced speech candidate y and the input speech x, and determines a driving excitation vector in the frame. The excitation code n2 determined by the codebook search control unit 14 and the linear prediction parameter code n1, which is the output of the linear prediction parameter encoding unit 3, are sent to the code transmission unit 4 and stored in a storage device according to the form of use. Or sent to the receiving side via a communication path.

FIG. 2 shows a CEL corresponding to the above encoding method.
4 shows a configuration example of a P decoding unit 20. Among the codes received from the transmission path or the storage medium, the linear prediction parameter code n1 is decoded into a synthesis filter coefficient by the linear prediction parameter decoding unit 21, and is sent to the synthesis filter 22 and, if necessary, the post-processing unit 30. The driving excitation code n2 is sent to the driving excitation vector generation unit 25, and an excitation vector corresponding to the code is generated. The configuration of the drive sound source vector generation unit 25 is as follows.
Driving excitation vector generation section 5 of encoding section 1 shown in FIG.
Is a configuration corresponding to. The synthesis filter 22 reproduces the sound s' using the driving sound source vector as an input. Post-processing unit 30
Is also called a post-filter, and performs processing to reduce the noise perception of the reproduced sound.

FIG. 3 shows a configuration example of a post filter. Post filters generally include enhancement of the spectral envelope,
Performs pitch enhancement using a comb filter. In FIG. 3, the decoded speech signal is converted to an MA having the inverse characteristic of the spectrum envelope.
The sound source waveform is extracted through a (moving average) type filter 32, and the periodicity of the pitch is enhanced through a comb type filter 38 having a tap at the pitch length position, and finally, the AR (self Through a (regression) type filter 39, an audio signal which is improved in auditory sense is obtained. Comb filters for emphasizing the periodicity of pitch may be realized by MA type or AR type, and the filter characteristics when the pitch period is an integer value are expressed by the following equations. Become like Here, a and bi are constants, and t is a pitch period. In the case of the MA type: H (z) = 1 + aZ- ^{t In} the case of the AR type: H (z) = 1 / (1-aZ- ^t ) In practice, the pitch period is often not an integer value. using, if the MA type: H (z) = 1 + aΣ i biZ -t + i for AR type: H be in the form of (z) = 1 / (1 -aΣ i biZ -t + i) Many.

In the above equation, the constants a, The pitch periodicity is indicated by bi.

[0010]

The problem with the post-filter in the CELP system is that pitch enhancement is performed on a frame-by-frame basis in the same manner as in encoding / decoding. Inevitably). That is, since the processing is performed on the assumption that the acoustic characteristics of the signal are constant within the frame, when the frame length is somewhat long (for example,
It is assumed that the pitch period and the degree of the periodicity of the pitch are constant in the frame in the portion of the transient characteristic in which the characteristic of the pitch period or the pitch periodicity changes in the frame. There is a problem that sufficient quality cannot be obtained in the processing.

[0011] It is an object of the present invention to provide a higher quality reproduced sound in a field of application of sound encoding / decoding without breaking a processing framework of a frame unit.

[0012]

According to a first aspect of the present invention, there is provided an audio signal post-processing method, comprising the steps of: storing an input audio signal for each frame in a storage unit; Passing the audio signal of the current frame and the past frame stored in the storage means through a linear filter using a linear prediction coefficient indicating an inverse characteristic of the spectral envelope of the audio signal to obtain a linear filter passing signal; Detecting a peak position of the waveform from the signal passed through the linear filter and extracting an average one-pitch length waveform based on the peak position; Calculate the correlation,
A step of sequentially searching for the peak of the cross-correlation as a pitch reference position, and dividing the linear filter passing signal into an area corresponding to a waveform having an accurate one pitch length based on the pitch reference position; The comb filter coefficient is calculated based on the pitch period and the strength of the pitch periodicity, and the linear signal passing signal is passed through a comb filter using the comb filter coefficient for each of the regions, and an audio signal output is calculated. And a step of obtaining.

According to a second aspect of the present invention, in the post-processing method of the first aspect, the audio signal obtained for each area is multiplied by a window function, and is multiplied by a window function of the immediately preceding area. And a step of superimposing the audio signal. According to a third aspect of the present invention, in the post-processing method of the first aspect, the comb filter coefficient is calculated for each of the regions. According to a fourth aspect of the present invention, in the post-processing method of the audio signal, the storage means for storing the input audio signal for each frame, and the audio signals of the current frame and the past frame stored in the storage means are stored in the storage means. A filter that passes through a linear filter that uses a linear prediction coefficient that indicates the inverse characteristic of the spectral envelope of the audio signal to obtain a linearly filtered signal;
A peak position detection / signal waveform cutout unit for detecting a peak position of a waveform from the linear filter passing signal and extracting an average one-pitch length waveform based on the peak position; A pitch reference position search unit for calculating a cross-correlation with the linear filter passing signal and sequentially searching for the peak of the cross-correlation as a pitch reference position; A boundary determining unit for dividing the region into regions corresponding to the pitch length waveform, a pitch correlation value calculating unit that calculates a comb filter coefficient based on the pitch period and the strength of the pitch periodicity of the linear filter passing signal, An audio signal output is obtained by passing the linear filter passing signal for each of the regions through a comb filter using the comb filter coefficient. .

According to a fifth aspect of the present invention, in the post-processing apparatus for an audio signal according to the fourth aspect, a multiplying section for multiplying the audio signal obtained for each area by a window function and a window function of the immediately preceding area. And a superimposing unit for superimposing the obtained audio signal. According to a sixth aspect of the present invention, in the post-processing apparatus for an audio signal according to the fifth aspect, the pitch correlation value calculator calculates a comb filter coefficient for each of the regions.

According to a seventh aspect of the present invention, there is provided a method for storing an audio signal input for each frame in a storage means, and the step of converting the audio signals of the current frame and the past frame stored in the storage means into the audio signal. Passing the signal through a linear filter using a linear prediction coefficient showing the inverse characteristic of the spectral envelope to obtain a linear filter passing signal; detecting a peak position of the waveform from the linear filter passing signal; Extracting a waveform having a single pitch length, calculating a cross-correlation between the waveform having the average one pitch length and the signal passing through the linear filter, and sequentially searching for the peak of the cross-correlation as a pitch reference position; Based on the pitch reference position, the signal passed through the linear filter
The linear filter passing signal is divided into regions corresponding to a pitch-length waveform, and a comb filter coefficient is calculated based on the pitch period and the strength of the pitch periodicity of the linear filter passing signal. A program is recorded to execute a procedure for obtaining an audio signal output by passing a type filter coefficient through a comb filter.

According to an eighth aspect of the present invention, in the recording medium storing the program according to the seventh aspect, the audio signal obtained for each area is multiplied by a window function, and is multiplied by a window function of the immediately preceding area. And a step of superimposing the audio signal on the audio signal. According to a ninth aspect of the present invention, there is provided a recording medium storing the program according to the seventh aspect, further comprising a step of calculating the comb filter coefficient for each of the areas.

According to the present invention, by providing the above-described configuration, a predetermined time delay is allowed in the post-filter, and in the section of frame length + delay time,
A pitch-synchronous post-filtering for detecting a pitch position and performing emphasis processing in units of one pitch is realized. As a result, even when the pitch period fluctuates within one frame, pitch enhancement processing can be performed while detecting and responding to minute fluctuations in pitch.
With only a slight increase in delay time, it can be applied to the frame processing framework of the encoding and decoding parts.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below with reference to the drawings. FIG.
Shows a configuration example of a post filter according to the present invention. MA
The filter unit and the AR filter unit may be common, and processing inside the filter unit is a feature of the present invention. (As described below, MA
More preferably, the filter and the AR filter are changed in accordance with the processing of the present invention. In the present invention, a delay buffer 31 is provided in order to provide a fixed time delay.

FIGS. 5 to 7 are diagrams schematically showing the processing of each processing unit in FIG. For example, in FIG. 5, the input frame position is the signal s ′ for one frame output from the synthesis filter 22. 5 to 7 are waveforms e after passing through the MA filter 32. The frame that is actually subjected to post-filter processing is a section represented as a processing frame, and the time difference between the input frame and the processing frame is delayed. It is assumed that an average pitch length or a provisional pitch length (hereinafter, referred to as “average pitch length”) in the processing frame has been obtained. As the average pitch length, the periodic code of the adaptive codebook may be used, or the autocorrelation function of the synthesized signal may be calculated again to calculate the average pitch length. If the average pitch length is longer than the frame length, there is almost no effect of the present invention, so that the conventional post-filter processing may be used. The present invention is effective when the pitch period is shorter than the frame length.

Hereinafter, the pitch period is shorter than the frame length.
It is assumed that this is the case. First, the peak position in the frame
The detector 33 detects the peak position of the signal in the processing frame.
(The maximum point of the amplitude). For example, this
In FIG. ₀And (In the example of FIG. 5, the processing frame
It will be located near the right boundary of the team. Frame position
The relative positional relationship between the position and the signal waveform is random
The position of the peak position in the processing frame
Is optional each time. ) Next, the peak position P₀Based on
From the signal waveform, the signal waveform cutout unit 34
Cut out a waveform of equal pitch length. This is shown in FIG.
The cut-out position is set to a half
It is good to cut out from the area | region of each switch length. Also, as shown in FIG.
As described above, you can cross the border of the processing frame.
Since the right end of the force frame position cannot be
If the start position exceeds the right end of the input frame position,
The right end of the cutout position is the right end of the input frame position.
Next, the pitch reference position search unit 35
Shift the waveform to the left and right, and
Calculate the correlation, P₀Near a position about an average pitch away from
Search for the position where the cross-correlation is maximum, and
Determine the sub-position. Repeat this pitch reference position search process
Returns the processing frame and the input frame.
The pitch reference position is determined in the region that has been set. In the example of FIG.
Indicates that there are three pitch reference positions in the processing frame,
Two places outside the frame are decided. Next,
In the field determination unit 36, as shown in FIG.
The signal to an accurate one-pitch waveform based on the location
The boundaries of the area. How to determine the boundary point
If (P₁+ P_Two) / 2, (P₀+ P₁) / 2
The midpoint of the sub-position may be the boundary, but it may be slightly
To the right, for example, P_Two+ (P₁−P_Two) * 2/3,
A 2: 1 internal dividing point may be used as a boundary point. Less than the middle point
Generally, it is better to move to the right.
The waveform rises sharply and converges slowly.
Is often observed.

The pitch emphasis processing is performed by analyzing the periodicity of the pitch for each area. First, the pitch correlation value calculator
At 37, an accurate pitch period t1 and a pitch correlation value a1 are obtained using the cross-correlation between the region 1 and the past signal waveform. Next, for the area 2, an accurate pitch period t2 and a pitch correlation value a2 are obtained. The accurate pitch period and pitch correlation value obtained in this way can be calculated according to the change, unlike the conventional post-filter method, even if the signal changes transiently in the frame. .

Finally, using the pitch period and the pitch correlation value obtained for each area, the comb filter 38 as described above is used.
For each region to perform pitch enhancement processing. In the case of such processing, since the coefficient of the pitch emphasis filter changes for each small area, a discontinuous sound may be generated, which may cause quality deterioration. In order to prevent this problem, the areas determined by the area boundary determination unit 36 are overlapped little by little (for example, about 10 samples, about 1.25 msec), and the overlapping parts are added with triangular windows, and gradually added. It is preferable that the characteristics change.

In the examples of FIGS. 5 to 7, the MA filter and the AR
The filter was assumed to be the same as before, but after determining each region as shown in FIG. 7, the MA filter and the AR filter also
It is more preferable to re-analyze the coefficient for each region and apply a filter so as to emphasize the spectral envelope for each region. In addition,
In the above-described embodiment, in the code-driven linear prediction encoding and decoding, an example is described in which the present invention is applied to an audio decoding method for reproducing a high-quality audio signal from an encoded bit sequence. The processing target is not limited to the decoded audio signal, and the present invention can be applied to an audio signal that is unknown whether it is a decoded audio signal input for each frame.

Further, a system to which the present invention is applied is referred to as a CP.
Computer with U and memory, terminal device and CD-
ROM, a magnetic disk device, constituted by a machine-readable recording medium such as a semiconductor memory, a computer reads a program for executing a procedure of a post-processing method of an audio signal stored in the recording medium, controls the operation of the computer,
Each element in the above embodiment is realized.

[0025]

The post filter according to the present invention is 4 kbits.
The subjective quality evaluation was performed by applying to the decoded signal of the / s speech coding method. The frame length as a post-filter processing unit was set to 10 msec. The delay time was set to 5 msec. As a result, a remarkable improvement was observed in the quality of the voice of a woman in particular, as compared with the conventional post-filter method. More specifically, the part where vowels such as "in the south" continuously change in a short time has reduced the noise sensation, so that it can be heard clearly. It should be noted that a significant improvement was observed in the female voice and the male voice did not change much compared to the conventional method because the present invention is effective when the pitch period is sufficiently shorter than the frame length. . Although the frame length of 10 msec corresponds to 100 Hz, the pitch frequency of male voice is generally 100 Hz or less (the pitch length is longer than 10 msec) in many cases.
On the other hand, since the pitch frequency of the female voice is about 200 Hz to 400 Hz, as shown in the example of FIGS.
3 pitches from That is, unlike the male voice, the pitch length of the female voice is sufficiently shorter than the frame length, so that a great improvement in the subjective quality was obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration example of a CELP encoding unit.

FIG. 2 is a block diagram illustrating a configuration example of a CELP decoding unit.

FIG. 3 is a block diagram showing a configuration example of a post filter.

FIG. 4 is a block diagram of a post-filter processing unit according to the present invention.

FIG. 5 is a view for explaining detection of a peak position in a frame.

FIG. 6 is a diagram for explaining waveform extraction and search for a pitch reference position.

FIG. 7 is a view for explaining determination of an area boundary.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 CELP encoding part 2 Linear prediction analysis part 3 Linear prediction parameter encoding part 4 Code transmission part 5 Driven excitation vector generation part 6 Adaptive codebook 7 Fixed codebook 8,9 Multiplication part 10 Weight creation part 11 Addition part 12,22 Synthesis filter 13 Distortion calculation unit 14 Codebook search control unit 21 Linear prediction parameter decoding unit 25 Driving excitation vector generation unit 30 Post-processing unit 31 Buffer 32 MA filter 33 In-frame peak position detection unit 34 Signal waveform extraction unit 35 Pitch reference position search Section 36 area boundary determination section 37 pitch correlation value calculation section 38 comb filter 39 AR filter

Claims (9)

[Claims] 1. A process for storing an audio signal input for each frame in a storage means, and the inverse characteristics of the spectral envelope of the audio signal in the current frame and the past frame stored in the storage means. A step of obtaining a linear filter passing signal by passing through a linear filter using a linear prediction coefficient shown, detecting a peak position of a waveform from the linear filter passing signal, and temporarily averaging one pitch length or provisionally based on the peak position. 1 pitch length determined (hereinafter, “average 1 pitch length”
That. A) extracting a waveform; calculating a cross-correlation between the waveform having an average pitch length of 1 and the linear filter-passed signal; and sequentially searching for the peak of the cross-correlation as a pitch reference position; The linear filter passing signal is divided into regions corresponding to waveforms of an accurate one pitch length based on the above, and a comb filter coefficient is calculated based on the pitch period and the pitch periodicity of the linear filter passing signal. Passing the linear filter-passed signal through the comb filter using the comb filter coefficient for each region to obtain an audio signal output. 2. The post-processing method for an audio signal according to claim 1, wherein the audio signal obtained for each area is multiplied by a window function, and the audio signal obtained by multiplying the window function of the immediately preceding area is superimposed. A post-processing method of the audio signal. 3. The post-processing method for an audio signal according to claim 1, wherein said comb filter coefficient is calculated for each of said regions. 4. A storage means for storing an audio signal input for each frame, and a linear signal representing the inverse characteristic of the spectrum envelope of the audio signal, wherein the audio signals of the current frame and the past frame stored in the storage means are linearly expressed. A filter that obtains a linear filter passing signal by passing through a linear filter using a prediction coefficient, a peak position detecting unit that detects a peak position of a waveform from the linear filter passing signal, and an average one pitch length based on the peak position or A signal waveform cutout unit for extracting a tentatively determined one-pitch length (hereinafter referred to as “average one-pitch length”) waveform; and calculating a cross-correlation between the average-one-pitch-length waveform and the linear-filter-passed signal. A pitch reference position searching unit for sequentially searching for the peak of the cross-correlation as a pitch reference position; An area boundary determining unit that divides the filter passing signal into an area corresponding to an accurate one-pitch length waveform, and calculates a comb filter coefficient based on a pitch period and a pitch periodicity of the linear filter passing signal. An audio signal post-processing device, comprising: a pitch correlation value calculation unit; and an audio signal output obtained by passing the linear filter passing signal for each of the regions through a comb filter using the comb filter coefficient. 5. The post-processing device for an audio signal according to claim 4, wherein: a multiplication unit that multiplies the audio signal obtained for each of the regions by a window function; and an audio signal obtained by multiplying a window function of the immediately preceding region. And a superimposing unit for superimposing the audio signal. 6. The post-processing apparatus according to claim 4, wherein said pitch correlation value calculation unit calculates a comb filter coefficient for each of said regions. 7. A first procedure for storing an audio signal input for each frame in a storage means, and a current frame and a past frame audio signal stored in the storage means are converted into a spectrum envelope of the audio signal. A second procedure of obtaining a linear filter passing signal by passing through a linear filter using a linear prediction coefficient indicating an inverse characteristic; detecting a peak position of a waveform from the linear filter passing signal; averaging 1 based on the peak position; Pitch length or provisionally determined one pitch length (hereinafter, “average one pitch length”
That. A) calculating a cross-correlation between the waveform having an average pitch length of 1 and the signal passing through the linear filter, and sequentially searching for the peak of the cross-correlation as a pitch reference position; The linear filter passing signal is divided into regions corresponding to waveforms of an accurate one pitch length based on a pitch reference position, and a comb filter coefficient is calculated based on a pitch period and a pitch periodicity of the linear filter passing signal. And a program for executing a fourth procedure of obtaining a sound signal output by passing the linear filter passing signal through the comb filter using the comb filter coefficient for each region. 8. A recording medium on which the program according to claim 7 is recorded, wherein the audio signal obtained for each area of the third procedure is multiplied by a window function, and the audio signal is obtained by multiplying the audio signal by a window function of the immediately preceding area. A recording medium recording a program, characterized by having a procedure for superimposing an audio signal. 9. A recording medium on which the program according to claim 7 is recorded, further comprising a step of calculating the comb filter coefficient in the third step for each of the areas. recoding media.