JPH07177031A - Voice coding control system - Google Patents

Abstract

PURPOSE:To improve reproducing sound quality by preventing foldover distortion by suppressing the change of each remainder pitch waveform in a frame, and sampling a representative remainder pitch waveform from a smoothed remainder pitch waveform. CONSTITUTION:A shift part 16 performs the phase matching of each remainder pitch, and a Fourier transformation part 17 performs the Fourier transformation of the remainder pitch waveform. Two-dimensional Fourier transformation for the remainder pitch waveform is performed by the transformation part 17 and a Fourier transformation part 19 at the next stage. Since a high-order component represents large change between the remainder pitch waveforms as a result of two-dimensional Fourier transformation, smoothing processing is performed by suppressing a component with such large change, and such processing is performed at a coefficient processing part 20. Thence, inverse Fourier transformation is performed at a two-dimensional inverse Fourier transformation part 21, and the smoothed remainder pitch waveform can be found. The latest remainder pitch waveform is sampled at a representative remainder pitch waveform sampling part 22 as the representative remainder pitch waveform of the frame, and it is quantized at a quantization part 23, then, it is sent to a reception side.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coding control system for highly efficiently coding a voice signal. In recent years, in enterprise communication systems, digital mobile radio communication systems, voice storage systems, etc., voice signals have been encoded with high efficiency,
It is attempting to improve the line utilization efficiency and reduce the storage capacity. In such high-efficiency encoding of audio signals, improvement of reproduced audio quality is desired.

[0002]

2. Description of the Related Art Various systems have already been proposed for high-efficiency coding of voice signals. For example, a method of transmitting a linear prediction coefficient obtained by linear prediction analysis and a parameter regarding a sound source is widely adopted.

FIG. 5 is an explanatory diagram of the linear predictive coding unit.
51 is a linear prediction analysis unit, 52 is a prediction coefficient quantization unit, 53
Is an inverse filter unit, 54 is a residual signal quantization unit, and 55 is a multiplexing unit.

The input speech signal is added to the linear prediction analysis unit 51 and the inverse filter unit 53, and in the linear prediction analysis unit 51, the prediction coefficient is obtained by the linear prediction analysis for each analysis frame, and the prediction coefficient quantization is performed. It is quantized in the part 52. The prediction residual signal is output from the inverse filter unit 53 to which the input audio signal is added, using this prediction coefficient as a coefficient, and is quantized in the residual signal quantization unit 54. Multiplexer 5
5 multiplexes the quantized prediction residual signal and the prediction coefficient, and sends them to the receiving side via the transmission path.

The decoding unit on the receiving side separates the multiplexed prediction residual signal and the prediction coefficient, adds the prediction residual signal to a synthesis filter having the prediction coefficient as a coefficient, and reproduces a voice signal. Will be done.

For efficient transmission of a prediction residual signal, a code-driven linear predictive coding method CELP for vector-quantizing a prediction residual signal of a fixed length and transmitting its index.
In addition, various methods such as a multi-pulse drive coding method MPC in which a prediction residual signal of a constant length is modeled by a finite number of pulse trains and an optimum pulse position and pulse amplitude are transmitted are known.

A representative waveform interpolation method (PWI; Prototype)
Waveform Interpolation) is known as a method in which the reproduced voice quality is less deteriorated even at a low bit rate of 4 kpbs or less. This method extracts a typical one-pitch waveform from the residual signal in the analysis frame of the input speech signal and quantizes it. The residual signal other than the residual pitch waveform is obtained by interpolating the representative residual pitch waveform in the previous frame and the representative residual pitch waveform in the current frame.

FIG. 6 is an explanatory view of the PWI system, in which a representative residual pitch waveform of the current frame having a length of L samples is represented by Z _n ,
The pitch cycle is shown as N _n , the representative residual pitch waveform of the previous frame is shown as Z _n−1 , and the pitch cycle is shown as N _n−1 , and one of the plural residual pitch waveforms in one frame is shown. The difference pitch waveform is quantized as a representative, and on the decoding side, the remaining residual pitch waveform P _n excluding the representative residual pitch waveform Z _n of the current frame is the representative residual pitch waveform Z of the previous frame. _It can be obtained by performing interpolation after cyclic matching of _n−1 with the representative residual pitch waveform Z _n of the current frame. That is, it can be obtained by the linear interpolation shown in the following equation. P (k) = (1-α (k)) Z _n-1 (k) + α (k) Z _n (k) (1) where 0 ≦ α (k) ≦ 1 P _n = P _n (0 ), P _n (1), ... P _n (N-1) Z _n = Z _n (0), Z _n (1), ... Z _n (N-1) Z _n-1 = Z _{n -1} (0), Z _n-1 (1), ... Z _n-1 (N-1)

[0009]

[Problems to be Solved by the Invention]
In the I method, the residual signal is obtained by interpolating the representative residual pitch waveform. However, if a continuous residual pitch waveform in a frame has a fluctuation component equal to or higher than the frame frequency, the residual residual waveform other than the representative residual pitch waveform is detected. In the interpolation process for generating the residual pitch waveform, aliasing distortion occurs, and the error from the actual residual signal becomes large. As a result, there is a problem that the quality of reproduced voice is deteriorated. It is an object of the present invention to prevent the occurrence of aliasing distortion and improve reproduced voice quality.

[0010]

The speech coding control system of the present invention will be described with reference to FIG. 1. The input speech signal is analyzed by the analysis unit 1 to obtain a linear prediction residual signal for each frame.
The representative residual pitch waveform extracting unit 2 extracts a representative residual pitch waveform in the frame, sends this representative residual pitch waveform as one piece of encoded information, and then outputs one representative residual of this encoded information. In the interpolation processing decoding unit 3 which has received the pitch waveform, the residual pitch waveform in the current frame is obtained by the interpolation processing using the representative residual pitch waveform of the current frame and the representative residual pitch waveform of the previous frame. In the speech coding control method, the smoothing unit 4 smoothes changes in each residual pitch waveform in the frame, and the representative residual pitch waveform extracting unit 2 extracts the representative residual pitch waveform. It is a thing.

The smoothing unit 4 performs a two-dimensional Fourier transform on the residual pitch waveform, smoothes the transform coefficient obtained by the two-dimensional Fourier transform, and then performs a two-dimensional inverse Fourier transform to obtain a representative residual pitch. The configuration may be added to the waveform extracting unit 2.

Further, the smoothing unit 4 can be configured to normalize the transform coefficient in the two-dimensional Fourier transform process.

[0013]

In the analyzing unit 1, the input speech signal is subjected to the linear prediction analysis for each predetermined frame to output the residual pitch waveform, and the smoothing unit 4 changes between the residual pitch waveforms in the frame. Is smoothed so that does not become large. Then, in the representative residual pitch waveform extracting section 2, one representative is selected from the residual pitch waveforms in the frame to make a representative residual pitch waveform, and this is sent as one of the encoded information of the input speech signal. . Therefore, in the interpolation processing decoding unit 3, interpolation processing is performed using the representative residual pitch waveform of the current frame and the representative residual pitch waveform of the previous frame to obtain another residual pitch waveform of the current frame. After that, when the audio signal is reproduced, aliasing distortion can be suppressed, so that the quality of reproduced audio can be improved.

Further, the smoothing section 4 performs a two-dimensional Fourier transform on the residual pitch waveform, and the high-order component of the transform coefficient indicates a portion where the change of the residual pitch waveform is large. For example, the value is set to zero, and the residual pitch waveform is restored by the two-dimensional inverse Fourier transform. That is, the residual pitch waveform can be smoothed.

Further, in the process of converting the residual pitch waveform, after Fourier transform is performed in the column direction, the amplitude of the transform coefficient is normalized to, for example, 1 and the normalized transform coefficient is Fourier-transformed in the row direction. Do the conversion. That is, the transform coefficients can be normalized in the two-dimensional Fourier transform process.

[0016]

EXAMPLE FIG. 2 is an explanatory view of an example of the present invention.
Is an encoding unit, 30 is a decoding unit, 11 is a voiced / unvoiced determination unit (V / UV), 12 is an inverse filter unit, 13 is a linear prediction analysis unit, 14 is a pitch analysis unit, and 15 is a residual pitch. Waveform forming unit, 16 is a shift unit, 17 is a Fourier transform unit (D
FT), 18 is an amplitude normalization unit (NOM), 19 is a Fourier transform unit (DFT), 20 is a coefficient processing unit (MOD), 2
1 is a two-dimensional inverse Fourier transform unit (2D IDFT), 22 is a representative residual pitch waveform extraction unit, 23 is a quantization unit, 31 is a pulse generation unit, 32 is a waveform shaping unit, 33 is a voiced / unvoiced switching unit, and 34 is A noise generation unit, 35 is an amplifier, 36 is a prediction synthesis filter unit, and 37 is an interpolation processing unit.

In the encoding section 10 on the transmission side, the input voice signal is judged for each analysis frame in the voiced / unvoiced judgment section 11 to determine whether it is a voiced sound or an unvoiced sound, and the judgment signal is determined. In addition to the multiplexing unit (not shown), it is sent to the receiving side as one piece of encoded information of the audio signal.

The linear prediction analysis unit 13 obtains a prediction coefficient from the input speech signal, and the pitch analysis unit 14 obtains the pitch period N. In the inverse filter unit 12, the prediction residual signal R is obtained by performing an inverse filter process on the input speech signal using the prediction coefficient from the linear prediction analysis unit 13.
Then, the residual pitch waveform forming section 15 performs a cutting process of the residual pitch waveform based on the pitch cycle N obtained by the pitch analyzing section 14. The configuration for performing such processing can be realized by a configuration similar to the configuration for obtaining the residual pitch waveform in the already known PWI method.

The shift unit 16 cyclically shifts the residual pitch waveform to match the phases of the residual pitch waveforms, and the Fourier transform unit 17 performs the Fourier transform of the residual pitch waveform.
The Fourier transform unit 17 and the next Fourier transform unit 19 perform two-dimensional Fourier transform processing on the residual pitch waveform, and in the process, the amplitude normalization unit 1
In step 8, normalization processing is performed.

As a result of the two-dimensional Fourier transform, the higher-order component shows a component that greatly changes between the residual pitch waveforms. Therefore, by suppressing the component having such a large change, smoothing processing is performed. Yes, the coefficient processing unit 20 performs the above-described processing, and then the two-dimensional inverse Fourier transform unit 2
In step 1, inverse Fourier transform processing is performed to obtain a smoothed residual pitch waveform. Then, the representative residual pitch waveform extracting section 22 extracts the temporally latest residual pitch waveform as a representative residual pitch waveform in this frame,
The quantizer 23 quantizes and quantizes and sends it to the receiving side. To the receiving side, this representative residual pitch waveform and pitch cycle N
And the voiced / unvoiced determination signal are multiplexed and transmitted.

In the decoding section 30 on the receiving side, pulses are generated from the pulse generation section 31 in accordance with the pitch cycle N,
In addition, the interpolation processing unit 37 performs interpolation processing using the representative residual pitch waveform. That is, by using the representative residual pitch waveform of the previous frame and the representative residual pitch waveform of the current frame, another residual pitch waveform of the current frame is obtained by interpolation processing and added to the waveform shaping section 32 to generate the voiced sound. Generate the waveform of.

Further, the voiced / unvoiced switching section 33 is switch-controlled by a decision signal of voiced sound and unvoiced sound by a route not shown in the figure. In this case, the waveform shaping section 32 is switched to. Further, the gain of the amplifier 35 is set in accordance with the frame power information by a path (not shown). Then, the amplified signal is added to the prediction synthesis filter unit 36 to reproduce the audio signal.

Since the prediction residual signal vector R from the inverse filter unit 12 can be regarded as a waveform in which each residual pitch waveform P _i is continuous, R = (P ₀ , P ₁ , P2, ... P _M-1 ) (2) P _i = (p (i, 0), p (i, 1), p (i, 2), ... P (i, N _i )) 0 ≦ i ≦ It can be expressed as (M-1). Note that N indicates the pitch period, and M indicates the number of residual pitch waveforms included in the frame.

FIG. 3 is an explanatory diagram of extraction of the residual pitch waveform. The pitch analysis unit 14 finds the pitch period N of the current frame from the input voice signal, and the pitch period N is traced back from the latest time of this current frame. The residual pitch waveform P _i is cut out from the prediction residual signal vector R at. In this case, the oldest residual pitch waveform P ₀ is shown to include a part of the waveform of the previous frame with respect to the temporally newest residual pitch waveform P ₂ .

Next, in the shift section 16, each residual pitch waveform P _i is cyclically shifted so that p (i, 0) becomes the maximum value, and the phases of each residual pitch waveform P _i are aligned. . Letting this be P _i again, the prediction residual signal vector R can be expressed in the form of an N × M matrix as shown by R = [] in the following equation (3).

The equation (3) is two-dimensionally Fourier transformed.
That is, the Fourier transform units 17 and 19 sequentially perform the Fourier transform in the column direction and the row direction. In this case, the Fourier transform unit 17 performs the Fourier transform in the column direction, the amplitude normalization unit 18 normalizes the amplitude of each transform coefficient to 1, and the next Fourier transform unit 19 normalizes the amplitude. Fourier transform is performed in the row direction on the transform coefficient. By this normalization, the quantization efficiency of the representative residual pitch waveform can be improved.

The transform coefficient c (i, j) obtained by the two-dimensional Fourier transform by the Fourier transform units 17 and 19 (however,
i = 0,1,2, ... N / 2, j = 0,1,2, ...
-M-1) can be represented in the form of a matrix of (N / 2) * M as shown as C = [] in the following equation (4).

In the voiced sound section, there is a high correlation between successive residual pitch waveforms. For example, in FIG.
When the two-dimensional Fourier transform of the residual pitch waveform was performed on the audio signal waveform of the frame shown in (2), the result shown in (b) was obtained. That is, the smaller the value of j of the conversion coefficient c (i, j), the larger the value of the conversion coefficient. The low-order component with respect to j is a component that changes slowly between the residual pitch waveforms, and the high-order component is a component that changes rapidly.

Therefore, the two-dimensional Fourier transform coefficient c (i, j)
Where the value of j is the threshold value K (M, N,
The smoothing process is performed with a coefficient larger than f _cut ) set to zero. That is, the coefficient processing unit 20 processes the conversion coefficient c (i, j). K (M, N, f _cut ) = MN f _cut / f _s (5) f _cut = f _F / 2 where f _cut is the cutoff frequency, f _s is the sampling frequency, and f _F is the frame frequency.

The _cut- off frequency f _cut is selected to be 1/2 of the frame frequency f _{F because} , in the interpolation process, aliasing distortion occurs for a fluctuation component larger than 1/2 of the frame frequency f _F. Therefore, the quality of reproduced voice is deteriorated. However, by smoothing as described above,
Since it is possible to avoid the occurrence of aliasing distortion in the interpolation processing, it is possible to improve the reproduced voice quality.

The two-dimensional Fourier transform coefficient smoothed by the coefficient processing unit 20 is inverse Fourier transformed by the two-dimensional inverse Fourier transform unit 21 to obtain a smoothed residual pitch waveform. Then, in the representative residual pitch waveform extracting unit 22, the temporally newest residual pitch waveform P
_M-1 is extracted as a representative residual pitch waveform in this frame, and quantized in the quantizer 23. And
As described above, the pitch period N and the voiced / unvoiced determination signal are quantized and multiplexed, respectively, and sent to the receiving side.

In the decoding section 30 on the receiving side,
Interpolation processing is performed by the interpolation processing unit 37 using the representative residual pitch waveform, and the residual pitch waveform in the frame is reproduced, so that the voiced sound can be reproduced.

[0033]

As described above, according to the present invention, a linear prediction residual signal is obtained for each frame in which an input speech signal is analyzed, a representative residual pitch waveform is extracted and transmitted, and the representative residual pitch waveform is extracted. In the voice coding control method for reproducing the residual pitch waveform in the current frame by interpolating the typical residual pitch waveform in the previous frame and the representative residual pitch waveform in the current frame. The smoothing unit 4 smoothes the change in the residual pitch waveform of, thereby preventing the occurrence of aliasing distortion in the case of interpolation processing,
There is an advantage that the reproduced voice quality can be improved.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is an explanatory diagram of an example of the present invention.

FIG. 3 is an explanatory diagram for extracting a residual pitch waveform.

FIG. 4 is an explanatory diagram of a two-dimensional Fourier transform coefficient.

FIG. 5 is an explanatory diagram of a linear predictive coding unit.

FIG. 6 is an explanatory diagram of a PWI method.

[Explanation of symbols]

 1 analysis unit 2 representative residual pitch waveform extraction unit 3 interpolation processing decoding unit 4 smoothing unit

Claims (3)

[Claims] 1. An input speech signal is obtained by an analyzing unit (1) for a linear prediction residual signal for each frame, and a representative residual pitch waveform extracting unit (2) obtains a representative residual pitch waveform within a frame. In the interpolation processing decoding unit (3) that has extracted and transmitted the representative residual pitch waveform as one piece of encoded information, and received one representative residual pitch waveform of the encoded information, In a speech coding control method for obtaining a residual pitch waveform in the current frame by an interpolation process using the representative residual pitch waveform and the representative residual pitch waveform of the previous frame, each residual pitch waveform in the frame The speech coding control method is characterized in that the smoothing section (4) smoothes the change of the above and the representative residual pitch waveform extracting section (2) extracts the representative residual pitch waveform. 2. The smoothing unit (4) performs a two-dimensional Fourier transform on the residual pitch waveform, smoothes the transform coefficient obtained by the two-dimensional Fourier transform, and then performs a two-dimensional inverse Fourier transform. , The representative residual pitch waveform extraction unit (2)
The voice coding control system according to claim 1, wherein the voice coding control system is added to the above. 3. The speech coding control system according to claim 2, wherein the smoothing unit (4) has a configuration for normalizing transform coefficients in the two-dimensional Fourier transform process.