JPH0731516B2 - Multi-pulse encoder - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a multi-pulse encoding device that compresses the bandwidth of a voice signal to improve the efficiency of voice communication, and particularly to quantization noise shaping (noise shaping). The present invention relates to a multi-pulse coding apparatus in which the pitch prediction of a speech signal and the pitch prediction of a speech signal are matched.

[Conventional technology]

Conventionally, this type of technology rationally applies quantization noise shaping, improves the signal-to-noise ratio perceptually, and uses linear prediction (LPC) coefficients calculated from speech signals to perform band compression. I'm thinking.

In general, a voice generation model is composed of a driving sound source caused by vibration of vocal cords and a vocal tract transmission characteristic. Impulse response with a synthesis filter that applies LPC coefficients is one of the effective ways to express vocal tract transmission characteristics, and the basic pitch of multipulse is considered as a way to express a driving sound source. Both of these enable band compression by the multi-pulse encoding device.

However, in the conventional multi-pulse coding, the pitch prediction coefficient for pitch prediction is directly obtained from the input speech signal, and the result and the multi-pulse coding premised on the quantization noise shaping do not always match well. It wasn't something to do.

Therefore, the multi-pulse search needs to be an efficient method.

[Problems to be solved by the invention]

As described above, the problem to be solved by the present invention is that the pitch prediction coefficient is directly obtained from the speech signal, so that it does not necessarily match with the multi-pulse search method based on the quantization noise shaping. There is a point that cannot be taken.

Therefore, it is an object of the present invention to provide a multi-pulse encoding device which solves the above drawbacks.

[Means for solving problems]

A multi-pulse coding apparatus of the present invention is a multi-pulse coding apparatus that codes an input speech signal and outputs a coded signal, in which the speech signal is input and a transfer function represented by the following formula: (Where P is the order of the linear prediction coefficient, αi is the prediction coefficient of the i-th sample, and ri is the attenuation coefficient of the i-th sample.) And a spectrum modification means for outputting a spectrum modified speech signal whose spectrum is modified by Waveform encoding means for performing an average amplitude difference function process on the modified speech signal, performing pitch prediction analysis according to the minimum value, and outputting a multi-pulse encoded signal.

〔Example〕

Next, the present invention will be described in detail with reference to the drawings illustrating an embodiment.

FIG. 1 is a block diagram showing a configuration of a first embodiment including the present invention, FIG. 2 is a block diagram showing a configuration of a second embodiment including the present invention, and FIG. 3 is a first and a second. Block diagram showing a configuration of a pitch prediction filter unit used in the embodiment,
FIG. 5 and FIG. 5 are charts showing the waveforms of the first and second embodiments, respectively.

First, the outline of the present invention will be described.

In the present invention, the prediction coefficient αi is obtained for the sampled input signal through linear prediction analysis (LPC). Next, from the preset damping coefficient ri (i = 1, ..., P, ri is generally empirically determined between 0 and 1) and this linear prediction coefficient αi The input voice signal is spectrally transformed by the filter represented by, and a spectrally transformed voice signal is obtained.
Further, the pitch period is obtained from the interval corresponding to the time of the minimum value of the signal through the average amplitude difference function (AMDF) processing. In this average amplitude difference function processing, a signal delayed by the interval of each sample is created with respect to the input one frame of the spectrum-modified voice signal, and the absolute value of the difference in amplitude between the delayed signal and the undelayed signal is calculated. Then, the average value for one frame is taken and the minimum value is searched. The time corresponding to the interval of the minimum values thus obtained is the pitch period.

Next, the cross-correlation coefficient between the voice signal and the impulse response signal is calculated in one frame, and the maximum amplitude position and
The positions of the delay zero of the impulse response signals arranged for each of the above pitch periods are matched, and the amplitudes of the two are made the same, and the amplitude and the position at this time are output as a part of the multi-pulse signal. After that, the same method is repeated for the remaining portion obtained by subtracting the latter from the former, and the conventional multi-pulse search is performed, and the amplitudes and positions at that time are output in descending order of amplitude. The multi-pulse search is performed by the above method.

Therefore, the pitch prediction coefficient calculated from the spectrum-modified speech is used.Therefore, the pitch prediction coefficient that is compatible with the process of subtracting the cross-correlation by autocorrelation when determining the multipulse is calculated by AMDF. There is.

In this way, the redundancy of the pitch signal is removed and the band compressibility is improved.

Next, a first embodiment of the present invention will be described with reference to FIGS. 1 and 4 focusing on the configuration and operation thereof. As shown in FIG. 1, the first embodiment including the present invention comprises an LPC analysis unit 1, a spectrum modifying unit 2, and a waveform coding unit 3 to form a multi-pulse coding unit, and a waveform decoding unit. 4 and
The LPC synthesizing filter unit 5 is provided to form a multi-pulse decoding unit. Hereinafter, an example of the internal configuration of each of the constituent elements will be shown and the respective elements will be described.

The sampled input speech signal 100 is input to the LPC analysis section 1 and the spectrum transforming means 2, and the LPC analysis section 1 analyzes the input speech signal to obtain a partial autocorrelation coefficient signal 101 up to the maximum degree P. K quantization / decoding unit 5 of spectrum transforming means 2
Output to 1.

The spectrum transforming means 2 includes a K quantizing / decoding unit 51, K ·
α conversion unit 52, α · r calculation unit 53, spectrum transformation unit 54
It has and. The partial autocorrelation coefficient signal 101 is input to the K quantization / decoding unit 51, and the quantized partial self-phase relation signal 104 is output to the waveform coding means 3 and is also decoded and K / α conversion unit. Output to 52.

In the K / α conversion unit 52, the decoded partial autocorrelation coefficient signal is input, and the signal representing the prediction coefficient αi is output to the α / r calculation unit 53 and the spectrum conversion unit 54 for each sample. The α · r calculation unit 53 calculates the product of the perceptual weighting coefficient ri selected empirically in advance and the prediction coefficient αi,
The signal representing αi · ri, that is, the α · r signal 103 is output to the spectrum modifying section 54 and the waveform coding means 3. The spectrum transformation unit 54 receives the α · r signal 103 and the linear prediction coefficient αi, and the transfer function is (Z is the sampled input audio signal, and H (Z) is the output of the filter).

Next, the waveform encoding means 3 will be described. The waveform encoding means 3 includes an AMDF calculation unit 55, a minimum value search unit 56, an impulse response unit 57, a cross correlation unit 58, an autocorrelation unit 59, a multipulse search unit 60, and a pulse quantization unit 61. And a multiplexing / combining unit 62.

First, the voice signal 102 is input from the spectrum modifying unit 2 to the AMDF calculation unit 55 and the cross-correlation unit 58. The AMDF calculator 55 produces a delayed signal according to a predetermined delay amount for the input one-frame audio signal 102 as described above, and calculates the difference between the delayed sampling sequence and the undelayed signal. The absolute value is calculated for each sample. Next, the average of this absolute value for one frame is obtained, and the average value and the predetermined delay amount become the AMDF output signal 109, and the values when the delay amount is changed to the local minimum value search unit 56 are sequentially obtained. Is output. In the minimum value search unit 56, the AMDF is set according to the delay amount.
The minimum value of the output signal 109 is searched, the time length corresponding to the delay amount between the minimum values is regarded as the pitch cycle, and the pitch signal 110 indicating the pitch cycle is output to the multiplexing synthesis section 62 and the multi-pulse search section 60. .

Next, the generation of multi-pulses in the waveform encoding means 3 will be described. First, the α · r signal 103 input from the spectrum modifying unit 2 is input to the impulse response unit 57. The impulse response unit 57 includes an LPC synthesis filter unit 57A and an impulse generation unit 57C, and the α · r signal 103 is input to the LPC synthesis filter unit 57A, and the coefficient thereof is satisfied. Further, an impulse signal (see FIG. 4 (a)) is output from the impulse generator 57C for each frame time length, and the impulse response signal 115 (fourth time) of the LPC filter 57A is output.
Figure (b)) is required. Impulse response signal 115
Is applied to the cross-correlation unit 58 and the auto-correlation unit 59. The auto-correlation unit 59 obtains the auto-correlation signal (see FIG. 4 (c)) of the impulse response signal and sends it to the multi-pulse search unit 60. On the other hand, the impulse response signal 115
(See FIG. 4 (b)) and the audio signal 102 (see FIG. 4 (d)) are input, and as a result, the cross-correlation signal (see
The figure (e) is sent to the multi-pulse search unit 60.

In the multi-pulse search unit 60, the cross-correlation signal (see FIG. 4 (e)) and the autocorrelation signal 115 (see FIG. 4 (c)).
When the pitch signal 110 and the pitch signal 110 are input, the pulse signal corresponding to the maximum value of the cross-correlation signal is the multi-pulse signal 111.
Become part of. At this time, the maximum value of the cross-correlation signal and the auto-correlation signal whose position and the position thereof are the same are subtracted from the cross-correlation signal. At the same time, the autocorrelation signal arranged at a position corresponding to each pitch period with the position of the autocorrelation signal as a reference is subtracted from the cross-correlation signal. As a result, the pulse signal corresponding to the maximum value of the subtracted cross-correlation signal becomes a part (here, the second pulse) of the multi-pulse signal. Thereafter, the method is continued until the autocorrelation signals are successively subtracted by the method as described above and a predetermined number of multipulses are obtained.

The multi-pulse signal 111 (see FIG. 4 (f)) thus obtained is output to the pulse quantizer 61.

In the pulse quantizer 61, the input multi-pulse signal 111
Is quantized, and a multi-pulse signal 112 indicating the position and amplitude of the multi-pulse and a maximum amplitude signal 113 indicating the maximum value of the amplitude of the multi-pulse are output to the multiplexing / combining unit 62.

In the multiplexing and combining unit 62, the partial autocorrelation coefficient signal 104, the pitch signal 110, the multi-pulse signal 112, and the maximum amplitude signal
113 is input, and a multiplexed signal 105 obtained by multiplexing these signals is output and input to the demultiplexing unit 63 of the waveform decoding means 4 through the transmission line.

Next, the waveform decoding means 4 will be described. The waveform decoding means 4 includes a demultiplexing unit 63, a pulse decoding unit 64, a pitch prediction filter unit 65, and a K decoding unit 66. Multiplexed signal from the multiplexing / synthesizing unit 62 of the waveform coding means 3
When 105 is input to the demultiplexing unit 63, the demultiplexing unit 63
Then, a partial autocorrelation coefficient signal 114 corresponding to the partial autocorrelation coefficient signal 104 and a pitch signal 12 corresponding to the pitch signal 110.
0 and a multi-pulse signal corresponding to the multi-pulse signal 112
122 and a maximum amplitude signal 123 corresponding to the maximum amplitude signal 113 are output.

Partial autocorrelation coefficient signal 11 output from demultiplexing unit 63
4 is a partial autocorrelation coefficient signal 1 decoded by the K decoding unit 66
It becomes 07 and is input to the LPC synthesis filter unit 5. Similarly, the multi-pulse signal 122 indicating the position and amplitude of the multi-pulse output from the demultiplexing unit 63 and the maximum amplitude signal 123 indicating the maximum value of the multi-pulse amplitude are decoded by the pulse decoding unit 64 and pitch prediction is performed. It is input to the filter unit 65. Further, the pitch signal 120 output from the demultiplexing unit 63 is also input to the pitch prediction filter unit 65. The multi-pulse signal 106 decoded from the pitch prediction filter is input to the LPC synthesis filter unit 5. The LPC synthesis filter unit 5 receives the multi-pulse signal 106 and outputs the partial autocorrelation coefficient signal 1
The output audio signal 108 is output under the control of 07.

Next, the second embodiment of the present invention will be described with reference to FIGS. Second
Referring to the figure, the second embodiment including the present invention is the LPC analysis unit 11
, A spectrum transforming means 12 and a waveform coding means 13 to form a multi-pulse coding portion, and a waveform decoding means.
The multi-pulse decoding section is configured by including 14 and the LPC synthesis filter section 15. An example of the internal configuration of each component will be shown below, and each component will be described.

The sampled input speech signal 200 is input to the LPC analysis unit 11 and the spectrum modification unit 12, is analyzed by the LPC analysis unit 11, and the partial autocorrelation coefficient signal 201 up to the maximum order P is the spectrum modification unit 12. Is output to.

The spectrum transforming means 12 includes a K quantization decoding unit 71, K.
α conversion unit 72, α · r calculation unit 73, and spectrum transformation unit 74
It has and. The sampled input speech signal 200 and the partial autocorrelation coefficient signal 201 are input to the spectrum transforming means 12, and the quantized partial autocorrelation coefficient signal 204 and the auditory-weighted spectrally transformed speech signal are input. The output of 202 and the waveform coding means 13 is the same as in the case of the first embodiment.

Next, the waveform encoding means 13 will be described. The waveform encoding means 13 includes an AMDF calculation unit 75, a minimum value search unit 76, an impulse response unit 77, a cross correlation unit 78, an autocorrelation unit 79, a multipulse search unit 80, and a pulse quantization unit 81. And a multiplexing / combining unit 82.

Of these, the voice signal 202 is input and the pitch signal 210 is output.
The operation of the AMDF calculation unit 75 and the local minimum value search unit 76, and the operation of the pulse quantization unit 81 which receives the multi-pulse signal 211 and outputs the multi-pulse signal 212 indicating the position and amplitude of the multi-pulse and the maximum amplitude signal 213, The operation of the multiplexing / combining unit 82 is similar to that of the first embodiment.

Next, the generation of multi-pulses in the waveform coding means 13 will be described. First, the α · r signal 203 input from the spectrum modifying means is input to the impulse response unit 77. The impulse response unit 77 includes an LPC synthesis filter unit 77A, a pitch prediction filter unit 77B, and an impulse generation unit 77C. The α · r signal 203 is input to the LPC synthesis filter unit 77A and its coefficient is determined. Also, the impulse generator 77
The impulse signal from C (see FIG. 5A) is output to the pitch prediction filter unit 77B, for example, every one frame time length.

The pitch prediction filter 77B has a variable delay element 9 as shown in FIG.
0A to 90N and an adder 91 are provided so that signals from the ends and connection points of the cascaded variable delay elements 90A to 90N are output to the adder 91. Also variable delay element 90A
.About.90N is controlled by the pitch signal 303 and has the same delay time as the pitch interval. Therefore, when the input signal 301 is input from one end where the variable delay elements 90A to 90N are cascaded, a signal having the same waveform as the input signal 301 is input to the variable delay elements at the pitch interval indicated by the pitch signal 303. Only the signal added to the input signal 301 is output. (Note that this pitch prediction filter is also used in the waveform decoding means in the first and second embodiments.) Therefore, when the pitch signal 210 is input to the pitch prediction filter unit 77B, the impulse generation unit 77C An impulse-shaped pitch signal (see FIG. 5 (b)) is generated at every pitch period interval of the pitch signal 210 with respect to the impulse signal input from
It is output to the LPC synthesis filter unit 57A. From the LPC synthesis filter unit 77A, the impulse response signal (see FIG. 5 (c)) for the impulse-shaped pitch signal is output to the cross-correlation unit 78.
Is output to the autocorrelation unit 79, and the autocorrelation unit 79 obtains an autocorrelation signal (see FIG. 5D) and sends it to the multi-pulse search unit 80. On the other hand, the impulse response signal (see FIG. 5C) and the voice signal 202 (see FIG. 5E) are input to the cross-correlation unit 78, and as a result, the cross-correlation signal (see FIG. ) Reference) is sent to the multi-pulse search unit 80.

In the multi-pulse search unit 80, the cross-correlation signal (see FIG. 5 (f)) and the auto-correlation signal (see FIG. 5 (d)) are input, and the pulse signal corresponding to the maximum value of the cross-correlation signal is multi-pulse. It becomes a part of the signal 211 (see FIG. 5 (g)). At this time, the cross-correlation signal is subtracted from the auto-correlation signal whose maximum value is equal to the maximum value of the cross-correlation signal. As a result, the pulse signal corresponding to the maximum value of the subtracted cross-correlation signal becomes a part (here, the second pulse) of the multi-pulse signal. After that, the method is continued until the auto-cross-correlation signals are successively subtracted by the above-mentioned method and a predetermined number of multi-pulses are obtained.

The multi-pulse signal 211 thus obtained is output to the pulse quantizer 81.

The multiplexing / synthesizing unit 82 has a partial autocorrelation coefficient signal 204, a pitch signal 210, a multi-pulse signal 212, and a maximum amplitude signal.
And 213 are input, and a multiplexed signal 205 obtained by multiplexing these signals is output. The signal output here is input to the demultiplexing unit 83 of the waveform decoding means 14 through the transmission line, as in the first embodiment.

Next, the waveform decoding means 14 and below will be described. The waveform decoding means 14 includes a demultiplexing unit 83, a pulse decoding unit 84,
A pitch prediction filter unit 85 is provided, and the demultiplexing unit 83 has a partial autocorrelation coefficient signal 214 corresponding to the partial autocorrelation coefficient signal 204 and a pitch signal 220 corresponding to the pitch signal 210.
And the multi-pulse signal 2 corresponding to the multi-pulse signal 212
22 and the maximum amplitude signal 223 corresponding to the maximum amplitude signal 213 are output. Further, from the waveform decoding means 14, the multi-pulse signal 206 and the decoded partial correlation coefficient signal 207
And are output. These signals are further input to the LPC synthesis filter unit 15, and the output audio signal 208 is output from here. The above operation is similar to that of the first embodiment.

Therefore, in the embodiment of the multi-pulse encoding device described above, in encoding the sampled input voice signal, the multi-pulse signal is calculated from the voice signal subjected to the preset perceptual weighting, and the voice signal The feature is that the pitch frequency of is calculated from the interval of the minimum value of the AMDF calculation result.

〔The invention's effect〕

As described in detail above, the multi-pulse encoding device of the present invention performs a quantization noise shaping process to generate a multi-pulse signal from an input speech signal, performs AMDF calculation, and pitches from the interval of the minimum value. Since the pitch prediction coefficient of the speech signal to be multi-pulse is obtained by obtaining the frequency, there is an effect that the adaptability of the multi-pulse decision becomes extremely good.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a first embodiment including the present invention, FIG. 2 is a block diagram showing a configuration of a second embodiment including the present invention, and FIG. 3 is a first and a second. Block diagram showing a configuration of a pitch prediction filter unit used in the embodiment,
FIG. 5 and FIG. 5 are tables showing waveforms of the first and second embodiments, respectively. 1,11 …… LPC analysis unit, 2,12 …… Spectral transformation means, 3,1
3 ... Waveform coding means, 4,14 ... Waveform decoding means, 5,15
...... LPC synthesis filter section.

Claims (1)

[Claims] 1. A multi-pulse encoding device for encoding an input speech signal and outputting the encoded signal, wherein the speech signal is inputted and a transfer function is expressed by the following equation. (Where P is the order of the linear prediction coefficient, αi is the prediction coefficient of the i-th sample, and ri is the attenuation coefficient of the i-th sample.) And a spectrum modification means for outputting a spectrum modified speech signal whose spectrum is modified by , A waveform encoding means for performing average amplitude difference function processing of the modified speech signal, performing pitch prediction analysis according to the minimum value thereof, and outputting a multi-pulse encoded signal.