JPH05127699A - Code excitation linear prediction and encoding system - Google Patents

Abstract

PURPOSE:To obtain sufficient synthesized speech quality even when a transmission bit rate is low. CONSTITUTION:With Ia, Is, and Ig supplied from an auditory error calculating circuit 316, an adaptive excitation code book 305 outputs and supplies an optimum code vector eai to a multiplier 308, and a gain code book 306 outputs a pair of optimum excitation gains gak, gsk and supplies the gak to a multiplier 308 and the gsk to a multiplier 310. The multiplier 308 supplies a multiplication vector eaik to an adder 309. The multiplier 310 supplies a multiplication vector eslk to the adder 309. The adder 309 supplies the added excitation vector (e) to a multiplier 312, which multiplies it by an excitation gain sigma and supplies the excitation vector eg to a composing filter 313. The composing filter 313 filters the excitation vector eg with the LSP coefficient obtained by an LSP inverse quantizer 303 to obtain a synthesized speech.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a code-excited linear predictive coding system, which can be applied as a high-quality compression coding system for voice signals.

[0002]

2. Description of the Related Art A conventional code-excited linear predictive coding system is disclosed in, for example, Document 1. Reference 1: "1989, Proc.IEEE Int.Conf.on Acoustics, Spe
ech and Signal Processing, PP.65-pp.68, NSJayant a
nd JH Chen, "Speech Coding with Time-Varying Bit A
llocations to Excitation and LPC Parameters "] FIG. 2 is a functional block diagram of an example of a code excitation linear predictive encoder.

In FIG. 2, the input original speech vector S is
It is supplied to the analysis circuit 102, where PARCOR (PARti
al auto-COR relation: partial autocorrelation) analysis, etc., and supplies the quantized short-time prediction coefficient αjp to the multiplexing circuit 111.
The vocal tract prediction coefficient αjq is supplied to the short-time filter 103.

Next, the switches 108 and 113 are opened, and the adaptive excitation code table 104 selectively outputs the adaptive excitation code vector ea (i) (i = 1 to n) based on the supplied adaptive excitation code index Ia. .. The gain control circuit 119 generates an adaptive excitation gain control code βap from the original speech vector S and the adaptive excitation code vector ea (i) and supplies it to the multiplexing circuit 111. Further, the gain control circuit 119 outputs the adaptive excitation gain control signal βap to the multiplier 12
2 to the adaptive excitation code vector βea (i) to control the gain to obtain the adaptive excitation code vector βea (i).
(i) is supplied to the adder 110.

The adder 110 supplies the adaptive excitation code vector βea (i) as the excitation vector e (i) to the filter for a short time as it is, and the synthesized speech vector Sw is subtracted from the subtractor 10.
Supply to 9. The subtractor 109 subtracts the original speech vector S and the synthesized speech vector Sw in component units, and supplies the error vector er (i) to the auditory filter 105. The auditory filter 105 performs filtering and supplies the vector ew (i) to the auditory error calculation circuit 106. The auditory error calculation circuit 106 calculates the root mean square gi of each component unit of the vector ew (i).
Is calculated, and i with which gi is minimized is supplied to the adaptive excitation code table 104 and the multiplexing circuit 111 as the index Ia of the optimum adaptive excitation code. Next, the switch 113 is closed and the stochastic excitation code table 10
7, the stochastic excitation code vector es (i) is multiplied by the stochastic excitation gain control signal βsq generated by the gain control circuit 123 in the multiplier 126 to perform gain control, and the stochastic excitation code vector βes (i) is the adder 110
Supply to. At this time, the gain control circuit 123 supplies the stochastic excitation gain control code βsp to the multiplexing circuit 111. The adder 110 uses the stochastic excitation code vector βes (i)
And the adaptive excitation code vector βea are added, and the excitation vector e (i) is supplied to the short-time filter 103 in the same manner as described above, and thereafter, the subtracter 109, the auditory filter 105, and the auditory error are processed by the same processing method. The stochastic excitation code index Is is obtained by processing in the calculation circuit 106 and supplied to the multiplexing circuit 111 and the stochastic excitation code table 107.

Next, the switches 108 and 113 are closed,
Adaptive excitation codebook 1 by excitation vector e (i)
The contents of 04 are updated. The stochastic excitation code index Is, the adaptive excitation code index Ia, the stochastic excitation gain control code βsp, the adaptive excitation gain control code βap, and the quantized short-time prediction coefficient αjp, which are obtained as described above, are multiplexed to obtain a total code. It was outputting C.

[0007]

However, in the above conventional method, the PARC for the input audio signal is used.
When the number of quantized bits in the OR analysis becomes small, the reproduced synthesized speech is distorted and the like, and there is a problem that a good speech synthesis quality cannot be obtained such that sufficient S / N cannot be obtained. Therefore, the number of quantization bits cannot be reduced, which is an obstacle to reducing the transmission bit rate.

This is because, in the case of PARCOR analysis, for example, in order to reduce the amount of information, the frame interval is set to 20.
For example, linear interpolation is performed every 2.5 msec in order to extend the time to msec and smooth the change of each parameter during that period. However, this linear interpolation is
This does not necessarily apply to the R coefficient, and as a result, distortion is increased in the spectrum of the reproduced synthesized voice, resulting in deterioration of sound quality. It can be considered that this is because the PARCOR coefficient does not correspond to a single definite physical quantity and is a complex compound parameter.

Further, since the gains for the adaptive excitation codebook and the stochastic excitation codebook are controlled independently by the scalar quantized coefficients, when the number of quantized bits becomes small, distortion or the like occurs, and the reproduced synthesis is performed. Since the S / N ratio of voice cannot be sufficiently obtained, the number of quantization bits cannot be reduced, which is an obstacle for reducing the transmission bit rate.

Since the adaptive excitation gain control code and the stochastic excitation gain control code are independently generated and transmitted,
There is a problem that the transmission bit rate tends to be increased, and as the types of excitation source codebooks increase, the types of gain control codes also increase, and the transmission bit rate must be increased.

The present demand for such an encoding method is that the quality of synthesized speech reproduced at a low transmission bit rate can be obtained satisfactorily, but it has been difficult to realize due to the above problems.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a code-excited linear predictive coding system capable of obtaining sufficient synthesized speech quality even at a low transmission bit rate. Is to provide.

[0013]

In order to achieve the above object, the present invention has been improved by the following characteristic means and method in a code excitation linear predictive coding system using a plurality of types of excitation source information. .. That is, a line spectrum pair analysis means for performing line spectrum pair analysis on a voice signal to generate line spectrum pair analysis information used for generating a synthesized voice, and a plurality of sets of gain control information consisting of a plurality of gain control information are stored. With a gain control codebook
Gain control is performed on the plurality of types of excitation source information by a plurality of gain control information in the optimum gain control information set, and a plurality of types of gain-controlled excitation source information and the line spectrum pair analysis information are used. Then, the audio signal is encoded and decoded.

[0014]

According to the present invention, since the line spectrum pair analysis is performed on the input voice signal, the line spectrum pairs for the main formants correspond relatively well with little deviation from the formant frequency. Therefore, the linear interpolation characteristic is good, the spectrum distortion due to the linear interpolation is small, and the deterioration of the sound quality is small. Even if the coefficient of the line spectrum pair is quantized to a low bit number, the sound quality can be improved as compared with the conventional PARCOR analysis.

Further, it is not necessary to transmit the gain information for each of the plural kinds of excitation source information. That is, since the gain control information for each excitation source information is formed in a set, only the information designating this set needs to be transmitted, so that the transmission bit rate can be reduced as compared with the conventional case. Moreover,
By setting the set of gain control information to the optimum set of gain control information in advance, it is possible to efficiently obtain good sound quality.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A preferred embodiment of the code-excited linear predictive coding system according to the present invention will be described with reference to the drawings.

In this embodiment, a code-excited linear predictive encoder and its decoder for performing forward code-excited linear predictive encoding will be described as an example.

Another object of this embodiment is to realize a code-excited linear predictive encoder and a decoder thereof that can obtain excellent synthesized speech quality even at a low bit rate.

In order to realize this object, in this embodiment, the vocal tract prediction parameters are mainly used for LSP (Line Sp).
and a circuit for converting the signal into an quantized parameter and quantizing it, and dequantizing the LSP parameter for interpolation.
A circuit for converting into vocal tract prediction parameters, a VQ (vector quantization) gain codebook in which the gains of the adaptive excitation codebook and the stochastic excitation codebook are vector quantized, and a backward type excitation gain prediction circuit are provided. There is.

FIG. 1 is a functional block diagram of the code excitation linear predictive coding circuit of this embodiment. FIG. 3 shows a functional block diagram of the code-excited linear predictive decoder of this embodiment.

In FIG. 1, the code excitation linear predictive coding circuit of this embodiment includes a vocal tract analysis circuit 301, an LSP quantizer 302, an LSP dequantizer 303, a multiplexing circuit 304, and an adaptive excitation. Codebook 305, VQ gain codebook 306, stochastic excitation codebook 30
7, a multiplier 308, an adder 309, and a multiplier 310
A gain control circuit 311, a multiplier 312, a synthesis filter 313, a subtractor 314, and an auditory filter 315.
And an auditory error calculation circuit 316. Here, the vocal tract analysis circuit 301, the LSP quantizer 302,
The LSP dequantizer 303 forms an LSP analysis system.

Next, the operation of FIG. 1 will be described. The original speech vector S collected in frame units is the vocal tract analysis circuit 30.
1 and the vocal tract prediction parameter aj is obtained.
This vocal tract prediction parameter aj is calculated by the LSP quantizer 302.
, The vocal tract prediction parameter aj is converted into an LSP parameter and quantized, and the LSP code Iw is converted into an LSP parameter.
It is supplied to the inverse quantizer 303 and the multiplexing circuit 304. L
The SP dequantizer 303 dequantizes the LSP parameter from the code Iw, performs interpolation, and then inversely converts the LSP parameter into the vocal tract prediction parameter aqj, and supplies this to the synthesis filter 313.

On the other hand, the adaptive excitation codebook 305 multiplies the adaptive excitation code vector eai (i = 1 to n) by the multiplier 3
Supply to 08. Also, the stochastic excitation codebook 307
Supplies the stochastic excitation code vector esl (l = 1 to m) to the multiplier 310. Further, the VQ gain codebook 306 outputs a pair of excitation gains gak and gsk (k = 1 to p), and the excitation gain gak is supplied to the multiplier 308.
The excitation gain gsk is supplied to the multiplier 310. Then, the multiplier 308 outputs the adaptive excitation code vector eai.
Is multiplied by the excitation gain gak to obtain a vector eaik and supplied to the adder 309. The multiplier 310 multiplies the stochastic excitation code vector esl by the excitation gain gsk to obtain a vector eslk, and supplies the vector eslk to the adder 309.

The adder 309 adds the vector eaik and the vector eslk in component units to obtain the excitation vector e and supplies it to the multiplier 312. Gain control circuit 311
Is the multiplier 3 obtained by obtaining the excitation gain σ from the past optimum excitation.
Supply to 12. The multiplier 312 receives the excitation vector e
Is multiplied by the excitation gain σ to obtain a vector eg, which is supplied to the synthesis filter 313.

The synthesis filter 313 filters the vector eg using the supplied vocal tract prediction parameter aqj as, for example, a tap coefficient of the filter to obtain a synthesized speech vector Sw and supplies it to the subtractor 314. Subtractor 314
Is subtracting the original voice vector S and the synthesized voice vector Sw in component units to obtain an error vector er and supplies it to the auditory filter 315. The auditory filter 315 filters the error vector er and outputs the output vector ew.
Is supplied to the auditory error calculation circuit 316. The auditory error calculation circuit 316 obtains the root mean square of each component of the vector ew, and the combination of i, j, and k that minimizes this value is the optimum adaptive excitation code index Ia of each codebook,
The stochastic excitation code index Is and the excitation gain code index Ig are supplied to the multiplexing circuit 304, the adaptive excitation code index Ia is supplied to the adaptive excitation code book 305, and the stochastic excitation code index Is is stochastically excited. The excitation gain code index Ig is supplied to the codebook 307, and the excitation gain code index Ig is supplied to the VQ gain codebook 306.

The adaptive excitation codebook 305 outputs the optimum adaptive excitation code vector eao according to the adaptive excitation code index Ia and supplies it to the multiplier 308. Further, the stochastic excitation codebook 305 outputs the optimal stochastic excitation code vector eso according to the stochastic excitation code index Is and supplies it to the multiplier 310. Also, V
The Q gain codebook 306 outputs an optimum pair of excitation gains gao and gso according to the excitation gain code index Ig, the excitation gain gao is supplied to the multiplier 308, and the excitation gain gso is supplied to the multiplier 310.

The multiplier 308 multiplies the optimum adaptive excitation code vector eao by the optimum excitation gain gao to obtain eao.
Supply gao to the adder 309. Also, the multiplier 310
Multiplies the optimal stochastic excitation code vector eso by the optimal excitation gain gso, and supplies eso · gso to the adder 309. The adder 309 performs addition of eao · gao and eso · gso, supplies the added value to the adaptive excitation codebook 305 as eopt, and updates the content.

There are various methods for searching the optimum code index by each codebook, but there is no particular limitation. For example, the adaptive excitation code vector is ea1, the stochastic excitation code vector is es1,
It is also possible to find the optimum adaptive excitation code vector, stochastic excitation code vector, and excitation gain by changing the excitation gain in order from (ga1, gs1) to (gap, gsp).

The multiplexing circuit 304 multiplexes the LSP code Iw thus obtained, the adaptive excitation code index Ia, the stochastic excitation code index Is, and the excitation gain code index Ig, and totals them. The code C is output and supplied to the decoder.

FIG. 3 shows a functional block diagram of the code-excited linear predictive decoder of this embodiment.

In FIG. 3, the code-excited linear predictive decoding circuit of this embodiment includes a demultiplexing circuit 402, an adaptive excitation codebook 403, and a VQ gain codebook 404.
A statistical excitation codebook 405, a multiplier 406,
Multiplier 407, adder 408, and gain control circuit 40
9, a multiplier 410, an LSP dequantizer 411, and a synthesis filter 412.

Next, the operation of FIG. 3 will be described. The input total code C is separated into each signal in the demultiplexing circuit 402, the LSP code Iw is supplied to the LSP dequantizer 411, the adaptive excitation code index Ia is supplied to the adaptive excitation codebook 403, The probabilistic excitation code index Is is the probabilistic excitation codebook 405.
And the excitation gain code index Ig is VQ.
It is supplied to the gain codebook 404.

The LSP dequantizer 411 dequantizes the LSP parameter from the LSP code Iw, further interpolates it, and then converts it into the vocal tract prediction parameter aj, and supplies it to the synthesis filter 412.

The adaptive excitation codebook 403 supplies the adaptive excitation code vector ea corresponding to the adaptive excitation code index Ia to the multiplier 406. The stochastic excitation codebook 405 also supplies the stochastic excitation code vector es corresponding to the stochastic excitation code index Is to the multiplier 407. The VQ gain codebook 404 is
The excitation gains ga and gs corresponding to the excitation gain code index Ig are output, and the excitation gain ga is calculated by the multiplier 40.
6 and the excitation gain gs is supplied to the multiplier 407.

The multiplier 406 calculates a vector eag obtained by multiplying the adaptive excitation code vector ea by the excitation gain ga.
Is supplied to the adder 408. Further, the multiplier 407 supplies the vector esg obtained by multiplying the stochastic excitation code vector es and the excitation gain gs to the adder 408. The adder 408 supplies the vector e obtained by adding the vector eag and the vector esg to the multiplier 410. Also,
This vector e is also supplied to the adaptive excitation codebook, and the content is updated by this vector e.

The gain control circuit 409 obtains the excitation gain σ from the past optimum excitation and supplies it to the multiplier 410. The multiplier 410 multiplies the excitation vector e by the excitation gain σ to obtain a vector eg and supplies the vector eg to the synthesis filter 412.

The synthesis filter 412 filters the vector eg using the supplied vocal tract prediction parameter aj as a filter coefficient, for example, to obtain the synthesized speech vector S and output it.

According to the above embodiment, the LSP parameter code Iw obtained by the LSP analysis is transmitted, and the decoder side LSP dequantizes the LSP code Iw to obtain the vocal tract prediction parameter aj, and the synthesis filter. Since the synthetic speech is reproduced in S., even if the number of quantization bits of the LSP parameter is reduced, the S / N of the reproduced synthetic speech can be improved as compared with the analysis by the conventional PARCOR method. Also, in the VQ gain code book, the optimum excitation gains gao and gso are determined by the excitation gain code index Ig.
Is output, and the adaptive excitation code vector is multiplied by the optimal excitation gain gao, and the stochastic excitation code vector is multiplied by the optimal excitation gain gso for encoding and decoding. Only the optimum excitation gain code index Ig is required. Conventionally, two systems of gain code indexes were transmitted. Therefore, sufficient voice quality can be obtained even at a low transmission bit rate.

Although the forward type code excitation linear predictive encoder and decoder have been described in the above embodiments, the present invention is not limited to this configuration. For example, a backward type code excitation linear predictive encoder and decoder. And can be applied to a decoder.

The VQ gain codebook stores two types of excitation gains as a pair, but the present invention is not limited to this, and a plurality of VQ gain codebooks are stored as a set according to the number of types of excitation codebooks to be gain controlled. It may be configured to.

[0041]

As described above, according to the present invention, a plurality of kinds of gain-controlled excitation source information and the line spectrum pair analysis information are provided, which are provided with the line spectrum pair analysis means and the gain control codebook. Since the voice signal is encoded and decoded by using, the sufficient synthesized voice quality can be obtained even at a low transmission bit rate.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a code excitation linear predictive encoder according to this embodiment.

FIG. 2 is a functional block diagram of a code-excited linear predictive encoder according to a conventional example.

FIG. 3 is a functional block diagram of a code excitation linear prediction decoder according to this embodiment.

[Explanation of symbols]

301 ... Vocal tract analysis circuit, 302 ... LSP quantizer, 30
3, 411 ... LSP dequantizer, 304 ... Multiplexing circuit,
305 ... Adaptive excitation codebook, 306 ... VQ gain codebook, 307 ... Stochastic excitation codebook, 30
8, 310, 312, 406, 407, 410 ... Multiplier, 309, 408 ... Adder, 311, 409 ... Gain control circuit, 313, 412 ... Synthesis filter, 314 ... Subtractor, 315 ... Auditory filter, 316 ... Auditory Error calculation circuit, 402 ... Demultiplexing circuit.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kenichiro Hosoda 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd.

Claims (1)

[Claims] 1. A code excitation linear predictive coding method using a plurality of types of excitation source information, wherein a line spectrum pair analysis is performed on a speech signal, and a line spectrum pair is generated which is used to generate synthetic speech. The gain control codebook for storing a plurality of sets of gain control information composed of a plurality of gain control information is provided, and the plurality of types of excitation are controlled by the plurality of gain control information in the optimum gain control information set. A code excitation line characterized by performing gain control on source information and performing encoding and decoding of a voice signal using a plurality of types of gain-controlled excitation source information and the line spectrum pair analysis information. Predictive coding method.