JPH08129400A - Voice coding system - Google Patents

Abstract

PURPOSE: To provide a voice coding system to improve the quality of a generated voice. CONSTITUTION: In a coder 10, a linear prediction analysis is performed against sound source signals by a linear prediction residual detecting means 11 for every frame and prediction coefficients are obtained. These coefficients are used against the signals and an inverse filter process is performed and a prediction residual signal R is obtained. A pitch analysis means 12 obtains a pitch period N corresponding to a maximum amplitude pitch of the sound source signals. A pitch waveform extracting means 13 extracts a residual pitch waveform Pi from the signal R in accordance with the period N. A two-dimensional Fourier transformation means 14 performs a two-dimensional Fourier transformation of the waveform Pi at the pitch period and transformation coefficients are obtained. A quantizing means 15 quantizes the transformation coefficients and quantized coefficients are obtained as sound source information. A two dimensional Fourier inversion transformation section 41 of a decoder 40 performs two-dimensional Fourier inverse transformation and sound source signal waveforms are obtained.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coding system for compressing a voice signal with high efficiency. In recent years, there has been a demand for highly efficient compression of voice signals in corporate communication systems, digital mobile radio systems, voice storage systems, and the like.

[0002]

2. Description of the Related Art An all-pole model obtained by linear prediction analysis is known to be a model with a good spectrum envelope of a speech signal, and in high efficiency coding of a speech signal,
A method of transmitting a linear prediction coefficient obtained by linear prediction analysis and a parameter related to a sound source is widely used.

FIG. 6 shows a block diagram of a speech coder using linear prediction analysis, and FIG. 7 shows a block diagram of a speech decoder, which will be described. The speech coder shown in FIG. 6 includes a linear prediction analysis unit 1, a prediction coefficient quantization unit 2, and an inverse filter 3.
And a residual signal quantizing unit 4 and a multiplexing unit 5, and the speech decoder shown in FIG.
And a synthesizing filter 8.

In the encoder shown in FIG. 6, the linear prediction analysis unit 1 performs a linear prediction analysis process for each analysis frame of an input speech signal to obtain a prediction coefficient, and the obtained prediction coefficient is quantized into a prediction coefficient. The part 2 quantizes.

Prediction residual signals obtained by passing the input speech signal through an inverse filter 3 using the quantized prediction coefficients as filter coefficients are quantized by a residual signal quantization unit 4,
The quantized prediction residual signal is multiplexed by the multiplexing unit 5 and transmitted to the transmission path.

In the decoder shown in FIG. 7, the quantized prediction residual signal transmitted from the encoder is separated into a prediction residual signal and a prediction coefficient by the separating unit 7, and the prediction coefficient is used as a filter coefficient. The encoded audio signal is reproduced by passing it through the synthesis filter 8.

At this time, some methods are known for efficiently transmitting the prediction residual signal. As the method, a code-driven linear predictive coding method (CEL) in which a predictive residual signal of a fixed length is vector-quantized and the index is transmitted.
P), a multi-pulse drive coding method (MPC) that models a prediction residual signal of a constant length with a finite number of pulse trains, and transmits an optimum pulse position and pulse amplitude.

In addition, a representative waveform interpolation method (Prototype Wavef
orm Interpolation (PWI) is a method that can obtain high quality at a low bit rate of about 4 Kbps or less. In this method, a typical one-pitch waveform is extracted from the residual signal in the voice signal frame, quantized and transmitted, and the other residual signals in the frame are compared with the representative residual pitch waveform in the previous frame. This is obtained by interpolating with the representative residual pitch waveform in the current frame.

[0009]

By the way, the above-mentioned C
In ELP and MPC, encoding is performed on the entire residual signal of a frame, so if the frame is lengthened to reduce the bit rate or the number of quantization bits is reduced, there is a problem that the voice quality deteriorates. there were.

On the other hand, in PWI, since only one pitch waveform in a frame is encoded and transmitted, it is possible to reduce the bit rate. However, the residual signal that was not transmitted on the decoding side is obtained by interpolation of the representative residual pitch waveform, so if there is a fluctuation component above the frame frequency in the continuous residual pitch waveform within the frame, it will be folded back by interpolation. Distortion occurs. For this reason, there is a problem that the error from the actual residual signal becomes large and the voice quality deteriorates.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a speech coding system capable of improving speech quality.

[0012]

FIG. 1 shows the principle of the speech coding system of the present invention. In the figure, 10 is an encoder. In this encoder 10, 11 is a linear prediction residual calculation means, which obtains a prediction coefficient by performing a linear prediction analysis on the audio signal for each frame, and uses the obtained prediction coefficient to perform an inverse filter process on the audio signal. The prediction residual signal R is obtained by performing.

Reference numeral 12 is a pitch analysis means for obtaining the pitch period N of the voice signal. Reference numeral 13 is a pitch waveform extracting means for extracting the residual pitch waveform Pi from the predicted residual signal R according to the pitch cycle N.

Numeral 14 is a two-dimensional Fourier transforming means, which obtains transform coefficients by performing two-dimensional Fourier transform on the residual pitch waveform Pi in pitch synchronization. Reference numeral 15 denotes a quantizing means, which obtains a quantized transform coefficient as sound source information by quantizing the transform coefficient.

Reference numeral 40 denotes a decoder, which comprises a two-dimensional inverse Fourier transform means 41 for obtaining a sound source signal waveform by performing a two-dimensional inverse Fourier transform.

[0016]

In the present invention described above, the vector (denoted as R) of the prediction residual signal R can be considered as a waveform in which the individual residual pitch waveforms Pi are continuous, and therefore can be expressed as follows.

R = (P ₀ , P ₁ , ..., P _M-1 ), and P _i = [p (i, 0), p (i, 1), ..., p (i, N _i )] (0 ≦ i ≦ M−1).

Here, M is the number of residual pitch waveforms Pi included in the frame. Each residual pitch waveform Pi can be extracted, for example, as shown in FIG. In this case, the temporally oldest residual pitch waveform P ₀ includes the waveform of the previous frame.

Next, each residual pitch waveform Pi is cyclically shifted so that p (i, 0) becomes the maximum value, thereby aligning the phases of each residual pitch waveform. Putting this again as Pi, the prediction residual signal vector R can be expressed in the form of the following N × M matrix.

[0020]

[Equation 1]

Next, this matrix is two-dimensionally Fourier transformed. That is, the Fourier transform is sequentially performed in the column direction and the row direction. The obtained complex two-dimensional Fourier transform coefficient c (i, j) is
It can be expressed by the following N / 2 × M matrix.

[0022]

[Equation 2]

Since there is a high correlation between the continuous pitch waveforms in the voiced sound section, the two-dimensional Fourier transform coefficient c (i, j) shown by reference numeral 16 in FIG. In the Fourier transform coefficient c (i, j), the value of the coefficient having a small value of j becomes large. However, FIG. 3 is a three-dimensional spectrum diagram showing the shape of c (i, j) and its transition.
FIG. 7 shows a voice signal waveform for which this c (i, j) is obtained.

The low-order component with respect to j is a component that changes slowly between pitch waveforms, and the high-order component is a component that changes rapidly. Therefore, the quantization precision is increased for a portion having a large amplitude indicated by reference numeral 18, that is, a coefficient having a small value of j, and a quantization is performed for a portion having a small amplitude indicated by reference numeral 19, that is, a coefficient having a large value of j. By lowering the quantization accuracy, it is possible to improve the quality of the reproduced signal as compared with the case where the quantization is performed uniformly using the same number of quantization bits. j
Coefficients larger than a certain value may not be transmitted at all.

The values of N and M are values that change for each frame. These two are in an inversely proportional relationship, and as the pitch period N increases, M decreases, and when N decreases, M increases.

When the pitch period N is large, the individual pitch waveform becomes long, but the change between consecutive pitch waveforms is small. On the other hand, when the pitch period N is small, the individual pitch waveform becomes short, but the change between continuous pitch waveforms becomes large.

According to the method of the present invention, the transform coefficient matrix shown in the above equation 2 is matrix-quantized and transmitted. Matrix quantization is performed by preparing one codebook corresponding to the value of N in a certain range. If the size of each codebook is kept the same, the distribution of the quantization precision to the phase information of the individual pitch waveform and the information of the change of the continuous pitch waveform is automatically optimized.

In the decoder 40 shown in FIG. 1, the transmitted quantized coefficient (quantized two-dimensional Fourier transform coefficient) is converted into 2
By performing the two-dimensional inverse Fourier transform again by the three-dimensional inverse Fourier transform means 41, the sound source signal waveform can be restored. A reproduced voice signal can be obtained by passing this sound source signal waveform through the predictive synthesis filter.

Therefore, in the method of the present invention, the deviation of the pitch-synchronized two-dimensional Fourier transform coefficient distribution of the residual signal is utilized, and the quantization accuracy is optimally distributed in accordance with the deviation, so that the encoded speech quality is improved. Can be improved.

[0030]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 4 is a block configuration diagram of a voice encoder according to the voice encoding system of an embodiment of the present invention, and FIG. 5 is a block configuration diagram of a voice decoder. In these figures, parts corresponding to parts of the principle diagram shown in FIG.

The speech coder 10 shown in FIG. 4 has a voiced / unvoiced determination section 21, a linear prediction analysis section 22 and an inverse filter 23.
A linear prediction residual calculation unit 11 and a pitch analysis unit 12
A pitch waveform extraction unit 13, a cyclic shift unit 24, a two-dimensional Fourier transform unit 14 including a first Fourier transform unit 25, an amplitude normalization unit 26, and a second Fourier transform unit 27,
The matrix quantizer 15 is provided.

The speech decoder 40 shown in FIG. 5 comprises a two-dimensional inverse Fourier transform unit 41, a noise generation unit 42, a changeover switch 43, an amplifier 44, and a prediction synthesis filter 45. There is.

In the encoder 10 shown in FIG. 4, the voiced / unvoiced decision unit 21 processes the input voice signal for voiced sound / voice for each frame.
The unvoiced sound is determined, the linear prediction analysis unit 22 performs linear prediction analysis to obtain a prediction coefficient, and the inverse filter 23 performs inverse filter processing on the input speech signal using the prediction coefficient to obtain a prediction residual signal. Find R. Further pitch analysis unit 1
In step 2, pitch analysis is performed to obtain a pitch period signal N corresponding to the voice signal.

Next, the pitch waveform extraction unit 13 extracts the residual pitch waveform signal Pi from the predicted residual signal R according to the pitch period signal N. Here, the residual pitch waveform signal Pi is output for each pitch period.

Next, the cyclic shift unit 24 brings the peak position of the residual pitch waveform signal Pi to the beginning so that the phases of the residual pitch waveform signals Pi are aligned. The two-dimensional Fourier transform coefficient is obtained by two-dimensional Fourier transforming the two-dimensional Fourier transform unit 14 on each of the residual pitch waveform signals Pi having the aligned phases.

In the two-dimensional Fourier transform, first, the vector representation (C of equation 2) of the residual pitch waveform signal Pi is Fourier-transformed in the column direction by the first Fourier transform unit 25, and then the amplitude normalization unit 26. The amplitude of each transform coefficient is normalized to 1,
The second Fourier transform unit 27 performs the Fourier transform in the row direction on the normalized transform coefficient. This processing can improve the quantization efficiency of the transform coefficient.

The transform coefficient thus obtained is matrix-quantized by the matrix quantizer 15 to obtain a quantized coefficient signal. When quantizing, a codebook corresponding to the pitch period value N is used. These codebooks are prepared in advance.

The quantized coefficient signal is transmitted to the decoder 40 shown in FIG. 5 together with the prediction coefficient, pitch period signal N, voiced / unvoiced decision result signal, and frame power signal. In the decoder 40 shown in FIG. 5, when the voiced / unvoiced determination result signal indicates voiced sound, the changeover switch 43 switches to the V side,
When it indicates unvoiced sound, it is switched to the UV side.

That is, in the case of voiced sound, the prediction residual signal obtained by subjecting the received quantized coefficient signal to the two-dimensional inverse Fourier transform in the two-dimensional inverse Fourier transform unit 41, that is, the residual pitch waveform signal, is changed over. Amplifier 4 through 43
4 is amplified and further passed through the prediction synthesis filter 45 to be converted into a voice signal.

On the other hand, in the case of unvoiced sound, white noise generated from the noise generator 42 is used as a sound source signal, and this sound source signal is passed through the predictive synthesis filter 45 to obtain a reproduced sound signal.

[0041]

As described above, according to the present invention,
There is an effect that the quality of the coded speech can be improved by optimally distributing the quantization accuracy according to the deviation of the distribution of the pitch-synchronized two-dimensional Fourier transform coefficient of the residual signal.

[Brief description of drawings]

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

FIG. 2 is a residual pitch waveform diagram for explaining the principle of the present invention.

FIG. 3 is a two-dimensional Fourier spectrum diagram of a voice signal waveform for explaining the principle of the present invention.

FIG. 4 is a block configuration diagram of a speech coder according to a speech coding system of an embodiment of the present invention.

[Fig. 5] Fig. 5 is a block configuration diagram of a voice decoder according to a voice encoding system of an embodiment of the present invention.

FIG. 6 is a block configuration diagram of a speech coder according to a conventional speech coding system.

[Fig. 7] Fig. 7 is a block configuration diagram of a speech decoder according to a conventional speech encoding method.

[Explanation of symbols]

 10 speech encoder 11 linear prediction residual calculation means 12 pitch analysis means 13 pitch waveform extraction means 14 two-dimensional Fourier transform means 15 quantization means 40 speech decoder 41 two-dimensional inverse Fourier transform means R prediction residual signal Pi residual pitch Waveform N pitch period

Claims (6)

[Claims] 1. A linear method for obtaining a prediction residual signal R by obtaining a prediction coefficient by performing a linear prediction analysis of a speech signal for each frame and performing an inverse filter process on the speech signal using the obtained prediction coefficient. Prediction residual detection means, pitch analysis means for obtaining the pitch period N of the voice signal, pitch waveform extraction means for extracting the residual pitch waveform Pi from the prediction residual signal R according to the pitch period N, A decoder provided with a two-dimensional Fourier transform means for performing a two-dimensional Fourier transform of the residual pitch waveform Pi in pitch synchronization to obtain a transform coefficient, and a quantizing means for quantizing the transform coefficient to obtain a quantized coefficient. A speech coding method characterized by having. 2. The speech according to claim 1, wherein when the quantizing means quantizes the transform coefficient, the quantized bit allocation is different depending on the amplitude distribution of each transform coefficient. Encoding method. 3. The speech encoding system according to claim 1, wherein said quantizing means matrix-quantizes said transform coefficient and transmits it. 4. The speech coding method according to claim 3, wherein, when the quantizing means performs the matrix quantization, weighting quantization is performed according to an amplitude distribution of the transform coefficient. 5. The speech coding system according to claim 1, wherein the two-dimensional Fourier transforming means normalizes the amplitude of the residual pitch waveform Pi when performing the two-dimensional Fourier transform. 6. The speech code according to claim 1, further comprising a decoder equipped with a two-dimensional inverse Fourier transform means for obtaining the excitation signal waveform by receiving the transform coefficient and performing a two-dimensional inverse Fourier transform. Method.