JP2946525B2 - Audio coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial application] The present invention relates to an audio signal having a low bit rate, particularly 4.8 kb /
The present invention relates to a speech coding system for performing high-quality coding with a relatively small amount of calculation in about s.

[Conventional technology]

As a method of encoding an audio signal at a low bit rate of about 4.8 kb / s, for example, Japanese Patent Application No. 59-272435 (Reference 1) and Japanese Patent Application No. 60-178911 (Reference 2) The pitch interpolation multipulse method described is known. According to this method, the transmitting side extracts, from the audio signal for each frame, a spectrum parameter representing a spectrum characteristic of the audio signal and a pitch parameter representing the pitch,
In the voiced section of the audio signal, one frame of the sound source signal is
One pitch section (representative section) of a plurality of pitch sections obtained by dividing a frame for each pitch section is represented by a small number of multipulses, and the amplitude, phase, spectrum, and pitch parameters of the multipulse in the representative section are transmitted. . In the unvoiced section, the sound source of one frame is represented by a small number of multipulses and a noise signal, and the amplitude and phase of the multipulse, and the gain and index of the noise signal are transmitted.

On the receiving side, in the voiced section, the amplitude and phase of the multi-pulses are interpolated using the multi-pulse of the representative section of the current frame and the multi-pulse of the representative section of the adjacent frame, and the pitch other than the representative section of the current frame is used. The multipulse of the section is restored to restore the driving sound source signal of the frame. In the unvoiced section, the driving sound source signal of the frame is restored using the multipulse and the index and gain of the noise signal. Further, the restored driving sound source signal is input to a synthesis filter using spectral parameters to output a synthesized voice signal.

[Problems to be solved by the invention]

According to the conventional method described above, in a voiced section, a sound source signal is represented by interpolating a small number of multipulses set in a representative section and a multipulse in a representative section of an adjacent frame. However, two types of transmission parameters, that is, the amplitude and the phase of the multi-pulse, are required. To encode these, a total of about 10 bits per pulse is required. Therefore, in order to apply to a bit rate of about 4.8 kb / s, Ozawa and Araseki et al.
speech coding with natural Speech quality "(ICASS
As described in P, pp, 457-460, 1986) (Literature 3) and the like, when the frame length is 20 ms, it is necessary to reduce the number of multi-pulses in a representative section to about four.
Therefore, with such a small number, the degree of approximation of the sound source signal in the representative section is not sufficient, and there is a problem that the sound quality is deteriorated especially in a male speaker having a long pitch cycle.

Further, in the conventional method, coefficients of a synthesis filter representing a spectral envelope characteristic of a speech signal are calculated using a linear prediction (LPC) analysis method. However, in the LPC analysis method, it is difficult to express the spectral envelope of the voice of a female voice with a short pitch period because the pitch is affected by the pitch of the female voice. The sound quality of the synthesized speech synthesized using the filter was degraded. This was remarkable in the region where the bit rate was low and the number of pulses was small, particularly in the region of 4.8 kb / s or less.

SUMMARY OF THE INVENTION It is an object of the present invention to solve the above-mentioned problems and to provide a speech encoding system with a good sound quality at about 4.8 kb / s with a relatively small amount of calculation.

[Means for solving the problem]

According to a first aspect of the present invention, a speech encoding method obtains a spectrum parameter representing a spectrum envelope and a pitch parameter representing a pitch for each frame of a predetermined time length from an input discrete speech signal, The audio signal is divided into pitch sections corresponding to the pitch period determined from the pitch parameter, and a sound source signal of one of the pitch sections is divided into a pulse and a codebook representing a spectral envelope characteristic of the sound source signal. Represents, the amplitude and phase of the pulse so as to reduce the error between the audio signal and the synthesized signal obtained by the pulse and the restored excitation signal obtained by the codebook and the spectrum parameter, and one from the codebook. A codeword is selected, and the spectral parameter is determined based on the restored sound source signal. Correct, the pitch parameter and the spectral parameter and the amplitude of the pulse, and outputs the information representing the phase and the code word.

According to a second aspect of the present invention, in the speech encoding method, a spectrum parameter representing a spectrum envelope and a pitch parameter representing a pitch are obtained for each frame of a predetermined time length from an input discrete speech signal, and the pitch parameter The audio signal of the frame is divided for each pitch section according to the pitch period obtained from the above, and a sound source signal of one pitch section of the pitch sections is divided into a pulse and a codebook representing a spectrum envelope characteristic of the sound source signal. In other pitch sections other than the pitch section, a correction coefficient for correcting the amplitude and phase of the pulse is obtained, and the correction coefficient is obtained from the pulse, the correction coefficient, a restored excitation signal obtained from the codebook, and the spectrum parameter. Determine the amplitude and phase of the pulse so as to reduce the error between the synthesized voice and the voice signal. Selecting one codeword from the codebook and modifying the spectral parameter based on the restored excitation signal, the pitch parameter, the spectral parameter, the amplitude and phase of the pulse, the correction coefficient, and the codeword. Is output.

[Action]

A first feature of the speech coding method according to the present invention is that, as shown in the block diagram of FIG. 3, in a voiced section, an excitation signal of a pitch section in a frame (normally about 20 ms) is given an amplitude and a phase. A pulse generator 700 that generates a small number of pulses, a codebook of filter coefficients representing the spectral envelope of the sound source signal, or a codebook 720 of the impulse response of the filter, and selecting one codeword from the codebook 720 This is to be expressed by a sound source signal forming unit 710 that forms a sound source signal. The synthesis filter 730 is driven by the sound source signal represented in this way to obtain a synthesized voice.

The second feature is that the spectrum parameters (hereinafter, filter coefficients) of the synthesis filter 730 are re-calculated using the sound source signal expressed as described above.

Now, as an example, the number of pulses of the pulse generator 700 is set to 1
And Further, the code book is composed of a set of impulse responses of a filter representing a spectral envelope of a sound source signal. This is defined as h _j (n) (j = 1−2 ^M ). This impulse response can be obtained by various methods. For example, an absolute value spectrum is obtained by performing a FFT (fast Fourier transform) on a predetermined number of samples for each frame of a prediction residual signal obtained by performing an LPC analysis on an audio signal, and calculating the absolute value spectrum.
FFT gives the impulse response. As another method, coefficients of a filter are obtained from the prediction residual signal by a well-known LPC analysis, and an impulse response of the filter is obtained. In addition to the above, known methods can be used. Codebooks are created in advance by training a large amount of audio data.

The selection of the pulse amplitude g, the phase m, and the code word h _j (n) from the code book is performed as follows. FIG. 4A shows an audio waveform of a certain frame. The frame is divided into pitch sections for each pitch period T of the pitch parameter obtained from the audio signal, and attention is paid to one pitch section (representative section) (FIG. 4 (b)). The audio signal in this section is x _k
(N). The selection of the optimal codeword from the pulse amplitude g, phase m, and codebook in this section is
This is performed so as to minimize the weighted error power represented by the following equation.
The weighted error power E _k in the representative section is It is represented by Here, _k (nm) = _ghj (nm) * _hs (n) (2) Here, w (n) indicates the impulse response of the auditory weighting filter. A specific configuration example is described in “A New Model of LPC Excitation for Producing Na
tural Sounding Speech at low Bit Rates ", Proc.ICASS
P, pp, 614-617, 1982, Reference 4). However, this filter may not be provided. _k (n)
Represents a sound source signal using a pulse and a j-th codeword selected from a codebook, and indicates a reproduced sound reproduced by passing the signal through a synthesis filter. H _s (n)
Indicates an impulse response of a synthesis filter for synthesizing speech. The symbol * indicates convolution integration. The following equation is obtained by substituting equation (2) into equation (1) and performing partial differentiation with g to set it to 0.

Here is the _{x wk = x k (n)} * w (n) x 'wk (n-m) = h j (n-m) * h s (n) * w (n) ...... (4) . The optimal set of g, m, h _j that minimizes the expression (1) is obtained as follows. Equation (3) is first calculated using a certain code word as the impulse response sequence h _j , and g and m are determined so as to minimize equation (1). This includes G, m that maximizes. Perform the above processing for all j, The set of g, m, j with the largest value of is the set to be determined.

With the above processing, the amplitude, phase, and code word of the pulse are obtained in the pitch section of interest. FIGS. 4 (c) and 4 (d) show the obtained pulse, the synthesized waveform x _k (n) obtained by driving the synthesis filter based on the obtained pulse and the excitation signal of the representative section generated by the selected code word, respectively. Show. The above processing may be performed for all pitch sections in the frame, or may be performed for only one pitch section (representative section).

Next, the recalculation of the coefficients of the synthesis filter will be described. The sound source signal of the representative section obtained by the pulse and the codebook as described above is defined as v (n).

v (n) = g · h _j (nm) (5) With the coefficient of the synthesis filter being a _i and the sound source signal v (n) obtained through the synthesis filter, And e (n) indicates an error signal. The coefficient a _i is determined so as to minimize the following equation.

Substituting equation (6) into equation (7) and partially differentiating the coefficients a _i to 0
The following equation is obtained.

Therefore, _ai can be obtained by solving equation (8). Here, the first term on the left side of equation (8) is the autocorrelation of x (n),
The term is the cross-correlation of v (n) and x (n). Equation (8) is solved by, for example, "Digital" by Rabiner, Schafer et al.
processing of speech signals "(Pren
tice-Hall 1978) (Reference 5).

〔Example〕

FIG. 1 shows a speech coding apparatus for implementing a speech coding method according to the first invention.

In FIG. 1, an audio signal is input from an input terminal 100, and an audio signal x (n) for one frame (for example, 20 ms) is stored in a buffer memory 110.

The spectrum parameter calculation circuit 140 performs a well-known LPC analysis on the linear prediction coefficient a _i from the audio signal of the frame as a spectral parameter representing the spectral characteristic of the audio signal of the frame, calculates a predetermined order M, and calculates an impulse. The response is output to the response calculation circuit 170 and the weighting circuit 200.

The pitch calculation circuit 130 calculates an average pitch period T as a pitch parameter from the audio signal of the frame. As this method, for example, a method based on an autocorrelation method is known. For details, refer to the pitch extraction circuits in the above-mentioned documents 1 and 2. In addition to this method, other well-known methods (eg, cepstrum method, SIFT method, modified correlation method, etc.)
Can be used.

The pitch encoding circuit 150 outputs a code obtained by quantizing the average pitch period T with a predetermined number of bits to the multiplexer 260, and outputs a decoded pitch period T ′ obtained by decoding the code to the pitch dividing circuit 205. , Sound source signal calculation circuit
Output to 220.

The codebook 175 includes a sequence h _j (n) (n = 1−1) of the impulse response of the filter representing the spectral envelope of the sound source signal.
L) sets (codebooks) of 2 ^M types are stored. Here, the code book is created by learning from impulse response data of a filter representing the spectral envelope of the residual signal, which has been analyzed in advance from a large number of predicted residual signals of the audio signal.
As a learning method, a learning method of vector quantization is known. For example, “Vector Quanti” by Makhoul et al.
zation in Speech Coding, "(Proc. IEEE, vol, 73, 11, 155
1-1588, 1985) (Reference 6).
Various known methods can be used to determine the characteristics of the filter representing the spectral envelope of the residual signal. For example, LPC analysis, covariance analysis, improved cepstrum analysis, or the like can be used for the residual signal.
For the LPC analysis and the covariance analysis, reference can be made to the aforementioned reference 5. For the improved cepstrum analysis, see “Improvement of spectral envelope by improved cepstrum method” by Imai et al. (Transactions of the Institute of Electronics, Information and Communication Engineers, J62-A, pp. 217-233, 1979)
Year) (Reference 7). Codebook 175, 2
^{The M} impulse response sequences h _j (n) (j = 1−2 ^M ) are extracted one by one from j = 1 to j = 2 ^M and output to the impulse response calculation circuit 170.

The impulse response calculation circuit 170 uses the linear prediction coefficient _ai from the spectrum parameter calculation circuit 140 to perform the impulse response h _w (n) of the synthesis filter weighted with the auditory sense.
And output h _j (n) from codebook 175
And the impulse response x ′ _wk (nm) obtained by performing the convolution calculation according to the equation (4) is converted to an autocorrelation function calculation circuit 180
Output to

The autocorrelation function calculation circuit 180 calculates the impulse response x ′ _wk
The autocorrelation function R _hh (n) of (nm) is calculated and output up to a predetermined delay time. Autocorrelation function calculation circuit
For the operation of 180, reference can be made to the above-mentioned documents 1 and 2.

The subtractor 190 subtracts the output of the synthesis filter 281 for one frame from the audio signal x (n) of the frame, and outputs the subtraction result to the weighting circuit 200.

The weighting circuit 200 passes the subtraction result through an audibility weighting filter whose impulse response is represented by _w (n), obtains a weighting signal _xw (n), and outputs it. The weighting method can be referred to the above-mentioned documents 1 and 2.

The pitch division circuit 205 divides the audio signal of the frame for each T ′ using the decoded pitch period T ′.

The cross-correlation function calculation circuit 210 calculates the weighted signal x _w (n)
And the impulse response x ′ _wk (nm) are input to calculate and output a cross-correlation function φ _{xh up} to a predetermined delay time. This calculation method can be referred to the above-mentioned references 1 and 2.

In the sound source signal calculation circuit 220, the representative 1
For one pitch section (representative section), in order to represent the sound source signal by the codebook h _j (n) and one pulse, the code word and the pulse amplitude g and phase m are obtained. Then g,
Equation (3) is used to calculate m. Then, as mentioned in the paragraph of the action repeats the outputs above processing from the codebook 175 for 2 ^M types as h _j (n),
A set of g, m, h _j (n) that minimizes the error power in the equation (1) is obtained by the method described in the section of operation. The code indicating the index of the selected codebook is multiplexer-multiplexed.
And outputs g and m to the encoder 230.

The encoder 230 encodes the amplitude g and the phase m of the pulse in the representative section with a predetermined number of bits and outputs the result. Furthermore, by encoding the information P ₁ number predetermined bit of which indicates the sub-frame location of the representative section outputs to the multiplexer 260. Further, these are decoded to obtain a drive signal restoring circuit 283,
Output to the parameter correction circuit 178.

The parameter correction circuit 178 generates the sound source signal v (n) in the representative section using the pulse amplitude and phase obtained in the representative section and the selected codeword. Further, using the audio signal x (n), a linear prediction coefficient according to the above equation (8)
a _i is recalculated, converted to K parameters, and output to the parameter encoding circuit 160.

The parameter encoding circuit 160 encodes the K parameter and outputs the code l _k to the multiplexer 260. The decoded value is converted to a linear prediction coefficient a _i ′ and output to the synthesis filter 281.

The drive signal restoration circuit 283 generates a sound source signal in the representative section using the pulse amplitude and phase obtained in the representative section and the selected codeword. In the other pitch sections, the amplitudes are linearly interpolated using the amplitudes of the pulses in the representative section of the preceding and succeeding frames to obtain pulses in the other pitch sections. Further, for the selected codeword, the codewords in the representative section are linearly interpolated to obtain an impulse response representing the spectral envelope of the sound source signal in another pitch section. The above process is performed by restoring the sound source signal of the frame.

The synthesis filter 281 receives the restored excitation signal and receives the linear prediction coefficient a _i ′ from the parameter encoding circuit 160.
To obtain a synthesized voice signal for one frame, and calculate an influence signal for the next frame for one frame,
This is output to the subtractor 190. The calculation method of the influence signal can be referred to Japanese Patent Application No. 57-231605 (Reference 8).

The multiplexer 260 outputs a combination of a code representing the amplitude and phase of the pulse in the representative section, a code representing the position of the representative section, a code representing the K parameter, a code representing the pitch period, and a code representing the selected codeword.

 Next, a second embodiment according to the present invention will be described.

FIG. 2 shows a speech encoding apparatus for implementing the speech encoding method according to the second invention. In the figure, components having the same reference numerals as those in FIG. 1 operate in the same manner as in FIG.

In FIG. 2, reference numeral 225 denotes an amplitude / phase correction calculation circuit. The amplitude / phase correction calculation circuit 225 calculates a correction coefficient for correcting the amplitude and phase of the pulse in the representative section in each pitch section other than the representative section in the same frame. Specifically, it is obtained as follows. The input voice, amplitude correction coefficient, and phase correction coefficient in the i-th pitch section are x _i (n), c _i , and d _i , respectively.

In the i-th pitch section, the amplitude and phase of the pulse in the representative section and the amplitude and phase of the sound source signal restored by the code word are corrected, and the reproduced signal _i (n) and the input audio signal x _i reproduced through a synthesis filter. The perceptual weighting error power with (n) can be written as follows.

Here, _i (n−T′−d _i ) = g · h (nm−T′−d _i ) * h _s (n) (10) The amplitude and phase correction coefficients c _i and d _i can be obtained so as to minimize the expression (10). (10) get by partially differentiating the amplitude correction coefficient c _i 0 Distant following equation Eq.

c _i = Σx _wi (n) _wi (n−T′−d _i ) / Σ _wi (n−T′−d _i ) _wi (n−T′−d _i ) (11) Various phase correction coefficients Calculate equation (11) for d _i ,
What is necessary is just to find the pair of c _i and d _i that maximizes the equation (11). The above processing is performed for all pitch sections other than the representative section in the frame, and the amplitude / phase correction coefficient of each section is encoded by the encoder 23.
Output to 0.

The drive signal restoration circuit 285 generates the sound source signal v (n) using the pulse amplitude and phase and the selected code word in the representative section of the frame. In the representative period other than the i-th pitch interval in the same frame, amplitude sound source signal v (n) representative period, the phase correction coefficient c _i, i-th pitch correction according to the following equation using the d _i Sound source signal d _{i for the} section
(N) is generated.

_{d i (n) = c i} · v (n-T'-d i) ...... (12) provided that v (n) = g · h j (n-m) ...... (13) where h _j (n ), G, m are the codebook codeword, pulse amplitude, and pulse phase.

The embodiments of the present invention have been described above. However, each of the embodiments described above is merely an example of the present invention, and various modifications thereof are conceivable.

For example, the pulse of the representative section may be recalculated in the sound source signal calculation circuit 220 using the linear prediction coefficient _ai obtained again in the parameter correction circuit 178. For this purpose, the recalculated linear prediction coefficients are passed through an impulse response calculation circuit 170 to recalculate the impulse response, and furthermore, the autocorrelation function calculation circuit 180 and the cross correlation function calculation circuit 210
The cross-correlation is recalculated, and these are calculated by the sound source signal calculation circuit 220.
And then recalculate the pulse. In addition, the pulse calculation, linear prediction coefficient correction, and pulse
It may be repeated a predetermined number of times. By adopting such a configuration, the amount of calculation increases, but the characteristics are improved.

The calculation of the pulse amplitude and phase and the selection of the code word may be performed not only in the representative section but also in all pitch sections in the frame. With such a configuration, the amount of information necessary for transmitting the sound source information increases, but the characteristics are improved.

The representative section may be fixedly determined in the frame, for example, at the center of the frame, or may be obtained by searching for a pitch section that minimizes the error between the synthesized speech and the input speech. Reference 1 can be referred to for the specific method of the latter.

Further, the number of pulses in the representative section may be two or more. This improves the characteristics, but increases the amount of transmitted information.

Further, with respect to a code word, linear interpolation may or may not be performed in a pitch section other than the representative section.

Further, although the codebook is the impulse response of the filter representing the spectral envelope of the prediction residual signal of the audio signal, it may be the coefficient of the filter. In such a configuration, it is necessary to convert a filter coefficient into an impulse response. Specifically, known coefficients such as a linear prediction coefficient, a K parameter, a logarithmic cross-sectional area ratio, a cepstrum, and a mel cepstrum can be used as the coefficients.

In the embodiment, linear prediction coefficients are encoded as spectral parameters, and LPC analysis is used as an analysis method. However, other well-known parameters such as LSP, LPC cepstrum, cepstrum, improved cepstrum, and generalized Cepstrum, mel cepstrum and the like can also be used. In addition, an optimal analysis method can be used for each parameter.

Further, in order to reduce the amount of calculation, calculation of the influence signal can be omitted. Thus, the driving signal restoring circuit
The 283, the synthesis filter 281, and the subtractor 190 become unnecessary, and the amount of calculation can be reduced, but the sound quality is reduced.

As is well known in the field of digital signal processing, the autocorrelation function is represented by a power spectrum on the frequency axis,
Since the cross-correlation function corresponds to the cross-power spectrum, it can be calculated from them. These calculations are described in “Digital Signal Pr
ocessing "(Prentice-Hall, 1975) (Reference 9).

〔The invention's effect〕

As described above, according to the present invention, a sound source signal (representative section) in one pitch section is represented by using a small number of pulses giving an amplitude and a phase and a codebook representing characteristics of the sound source signal. Since the spectral parameters are re-calculated using such a sound source signal, a great effect is obtained at a bit rate of about 4.8 kb / s, in which the approximation degree of the sound source signal is higher than in the conventional method and a good synthesized voice can be obtained. is there.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech encoding device for explaining an embodiment of a speech encoding system according to the first invention, and FIG. 2 is a diagram explaining an embodiment of a speech encoding system according to the second invention. FIG. 3 and FIG. 4 are diagrams for explaining the principle of the present invention. 110 buffer memory 130 pitch calculation circuit 140 spectrum parameter calculation circuit 150 pitch coding circuit 160 parameter coding circuit 170 impulse response calculation circuit 178 parameter correction circuit 175,350,720 codebook 180 ... Autocorrelation function calculation circuit 205 ... Pitch division circuit 210 ... Cross correlation function calculation circuit 220 ... Sound source signal calculation circuit 225 ... Amplitude / phase correction calculation circuit 230 ... Encoder 260 ... Multiplexer 281,360,730 ... Synthesis filter 283 …… Drive signal restoration circuit

────────────────────────────────────────────────── ─── Continued on the front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/20 H03M 7/30 H04B 14/04 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A spectrum parameter representing a spectrum envelope and a pitch parameter representing a pitch are obtained for each frame of a predetermined time length from an input discrete voice signal, and a voice signal of the frame is obtained from the pitch parameter. Divide every pitch section according to the obtained pitch cycle,
An excitation signal of one of the pitch sections is represented by a pulse and a codebook representing a spectral envelope characteristic of the excitation signal, and a restored excitation signal is generated from the pulse and the codeword. Generate a composite waveform from the parameters, determine the amplitude and phase of the pulse so as to reduce the error between the composite waveform and the audio signal,
Selecting one codeword from the codebook, correcting the spectral parameter based on the restored excitation signal, the pitch parameter, the spectral parameter, the amplitude and phase of the pulse, and information representing the codeword; A speech encoding method that outputs 2. A spectrum parameter representing a spectrum envelope and a pitch parameter representing a pitch are determined for each frame of a predetermined time length from an input discrete speech signal, and the pitch parameter is determined according to a pitch period determined from the pitch parameter. Divided the audio signal of the frame for each pitch section,
The excitation signal of one of the pitch sections is represented by a pulse and a codebook representing the spectral envelope characteristic of the excitation signal, and the amplitude and phase of the pulse are corrected in other pitch sections than the pitch section. A correction coefficient to be obtained is generated, a restored excitation signal is generated from the pulse, the correction coefficient, and the codeword, a composite waveform is generated from the restored excitation signal and the spectrum parameter, and an error between the composite waveform and the audio signal is reduced. To determine the amplitude and phase of the pulse, select one codeword from the codebook, modify the spectral parameters based on the restored excitation signal, the pitch parameter and the spectral parameters and the pulse amplitude , Phase, the correction coefficient, and information representing the codeword