JPH05173596A - Code excitation linear predicting and encoding method - Google Patents

Abstract

PURPOSE:To obtain superior synthesized speech quality even at a low bit rate by generating a specific impulse response according to a vocal cord parameter and performing convolution processing for a statistical excitation vector, and then adding it to an adaptive excitation vector and generating an excitation vector. CONSTITUTION:The statistical excitation vector esl is converted and corrected before an adder 115 adds the adaptive excitation vector eai outputted from an adaptive excitation code book 107 and the statistical excitation vector esl outputted from a statistical excitation code book 108 to generate the excitation vector (e). For the conversion correction, a code vector converting circuit 109 performs the convolution processing for impulse response showing the short-time characteristics (vocal cord parameter) or/and long-time characteristics (pitch lag) of an input speech vector. Then the frequency characteristics of the input speech vector S are included in the statistical excitation vector esl and low-bit encoding is performed to improve the synthesized speech quality.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a code excitation linear predictive coding apparatus.

[0002]

2. Description of the Related Art A code-excited linear predictive coding apparatus provided with an adaptive excitation codebook accumulating optimum excitation vectors for past original speech vectors and a statistical excitation codebook accumulating predetermined excitation vectors, For example,
It is disclosed in the following documents. Reference name: N. S. Jayant and J. H. Ch
en, "SpeechCoding with Tim
e-Varying Bit Allocations
to Excitation and LPC Pa
parameters ”, Proc. ICASSP, pp6
5-68,1989. In this paper, it is considered that the residual signal for speech synthesis is composed of a periodic signal and a noisy signal, and the statistical excitation codebook is made to correspond to the noisy signal. Corresponding to the periodic signal, and in order to select and determine the optimum excitation vector for the original speech vector, a plurality of excitation vectors are read based on the plurality of adaptive excitation vectors and the plurality of statistical excitation vectors read therefrom. Create a vector.

[0003]

SUMMARY OF THE INVENTION In the above documents,
The statistical excitation vector is predetermined, and if the number of vectors is reduced as the bit rate is reduced, there is a problem that fixed vectors cannot be used and the synthesized speech quality deteriorates. Therefore, it is an object of the present invention to provide a code-excited linear predictive coding apparatus that can obtain excellent synthesized speech quality even at a low bit rate.

[0004]

According to the present invention, a synthetic speech vector is created on the basis of a quantized vocal tract parameter and an excitation vector, and an error between the input speech vector and this synthetic speech vector is evaluated. , A code-excited linear predictive coding device that determines the index of the adaptive and statistical excitation codebook corresponding to the excitation vector most suitable for the input speech vector at the current time, and outputs the vocal tract parameter and both indexes as a total code. is there. Then, create a specific impulse response based on the vocal tract parameters, and using the impulse response,
The statistical excitation vector is provided with a unit for performing a convolution process and then adding the adaptive excitation vector to create the excitation vector. In the first invention, the impulse response corresponds to the impulse response of the transfer function expressed using the vocal tract parameter, and for example, the impulse of the transfer function H (Z) shown in the following equation (1). It corresponds to the response.

[0005]

[Equation 1]

However, A and B are constants of 0 <A <B <1, a _qj is an output of the LPC dequantizer 104, and p is a vocal tract analysis order.

In the second aspect of the invention, the impulse response corresponds to the impulse response of the transfer function H (Z) expressed by the following equation. H (Z) = 1 / (1-εZ ^−L ) (2) where ε is a constant 0 <ε ≦ 1, and L is a pitch lag calculated from the index of the adaptive excitation vector.

[0008]

In the present invention, the statistical excitation vector is converted and corrected before the adaptive excitation vector and the statistical excitation vector are added to create the excitation vector. The conversion correction is performed by convolving the impulse response representing the short-term property (vocal tract parameter) or the long-term property (pitch lag) of the input speech vector, or the impulse response representing both. This convolution process can be expressed by the following equation (3), where esl is the output statistical excitation vector of the statistical excitation codebook, escl is the converted statistical excitation vector, and h is the impulse response. escl = esl * h (3) where escl = [x ₀ , x _1, ..., x _n-1 ], es
l = [y ₀ , y _1, ..., y _n-1 ], h = [h ₀ ,
h _1, ..., H _n-1 ], [] are column vectors, x, y, h
Is each element, and n is the subframe length (or frame length).

The impulse response is the impulse response of the transfer function expressed by using the vocal tract parameter in the case of the first invention which uses the short-term property of the input speech vector, and the first response which uses the long-term property. In the case of the invention of No. 2, it is the impulse response of the transfer function expressed using the pitch lag.

Theoretically, the frequency characteristic of the input speech vector should be borne by the vocal tract parameter and the pitch lag, and the residual signal should have a flat characteristic in frequency. The residual signal (when passed through) has frequency characteristics. The present invention utilizes this (or vice versa) to improve the synthesized speech quality by adding the frequency characteristic of the input speech vector to the statistical excitation vector.

[0011]

1 is a block diagram showing a first embodiment of a code-excited linear predictive encoder according to the present invention. In FIG. 1, an input voice vector S, which is collected as a vector from a terminal 101 by a frame unit, is first input to a vocal tract analysis circuit 102 to calculate a vocal tract prediction parameter aj. The vocal tract analysis circuit 102 sends the vocal tract prediction parameter aj to the LPC quantizer 103. The LPC quantizer 103 quantizes the vocal tract prediction parameter aj, and its code Ic (LPC code) is quantized by the LPC dequantizer 10.
4. Send to the multiplexing circuit 106. LPC inverse quantizer 1
04 uses the LPC code Ic as the vocal tract prediction parameter aqj
To the synthesis filter 105. Next, the adaptive excitation codebook 107 uses the adaptive excitation vector eai.
(I = 1 to n), the statistical excitation codebook 108 outputs the statistical excitation vector esl (l = 1 to m), and the VQ gain codebook 110 outputs the excitation gains βk and γk (k = 1 to p), respectively.

The code vector conversion circuit 109 convolves the statistical excitation vector esl from the codebook 108 with the impulse response of the filter transfer function H (Z) represented by the following equation (4) to calculate the modified statistical excitation vector escl. To do.

[0013]

[Equation 2]

Here, a _qj is the output of the LPC dequantizer 104, and p is the vocal tract analysis order. Adaptive excitation vector eai
Is multiplied by the gain βk by the multiplier 113, and the vector e
aik, and the modified statistical excitation vector escl is multiplied by the gain γk by the multiplier 114 to obtain the vector esclk.
Becomes The adder 115 calculates the excitation vector e by adding the vector eaik and the vector esclk in component units. The synthesis filter 105 calculates a synthesized speech vector Sw for the excitation vector e and sends it to the subtractor 116. The subtractor 116 subtracts the synthesized speech vector Sw and the input speech vector S in component units to obtain an error vector e.
Send r to the perceptual filter 111.

The perceptual filter 111 calculates the perceptual error vector ew for the error vector er from the perceptual error calculation circuit 112.
To send to. The perceptual error calculation circuit 112 calculates the root mean square of each component of the perceptual error vector ew, and determines the excitation vector (that is, the combination of i, l, k) that minimizes this value as the optimum input speech vector at the current time. It is determined as an exciting vector. Then, the indices Ia, Is, and Ig of the respective codebooks at that time are set to the adaptive excitation codebook 10
7, statistical excitation codebook 108, VQ gain codebook 110, and multiplexing circuit 106.

The adaptive excitation codebook 107 is the optimal adaptive excitation vector eao according to the index Ia, and the statistical excitation codebook 107 is the optimal statistical excitation vector eso according to the index Is, and the VQ gain codebook 1
10 is the optimum VQ gain β according to the index Ig
It outputs o and γo, respectively. Code vector conversion circuit 10
9 transforms the optimum statistical excitation vector eso into the optimum modified statistical excitation vector esco. These eao, es
The optimum excitation vector eopt composed of co, βo, and γo is input to the adaptive excitation codebook 107, and the contents of the adaptive excitation codebook 107 are updated. The multiplexing circuit 106 transmits Ic, Ia, Is, and Ig as the total code C to the receiving side through the output terminal 113.

FIG. 2 shows a block diagram of a code-excited linear predictive decoder corresponding to the encoding of FIG. In FIG. 2, the total code C input from the input terminal 201 is separated by the demultiplexing circuit 212 into an LPC code Ic, an adaptive excitation code index Ia, a statistical excitation code index Is, and a VQ gain code index Ig,
LPC dequantizer 202 and adaptive excitation codebook 2 respectively
04, statistical excitation codebook 205, and VQ gain codebook 207. LPC dequantizer 202
Converts the LPC code Ic into a vocal tract prediction parameter aj and sends it to the synthesis filter 203. The adaptive excitation codebook 204 is the adaptive excitation vector ea corresponding to the index Ia, the statistical excitation codebook 206 is the statistical excitation vector es corresponding to the index Is, and the VQ gain codebook 207 is the excitation gain β and γ corresponding to the index Ig, respectively. Output.

The code vector conversion circuit 206 converts the vector es into the vector esc in the same manner as the encoder of the first embodiment. The excitation vector e is composed of ea, esc, β, and γ. The synthesis filter 203 calculates a synthesized speech vector S for the excitation vector e and outputs it from the output terminal 211. The contents of the adaptive excitation codebook 204 are updated by the vector e.

A second embodiment of the code-excited linear predictive encoder according to the present invention will be described with reference to FIG. The code vector conversion circuit 109 is the same as the encoder of the first embodiment except for the code vector conversion circuit 109.
The operation of will be described. The code vector conversion circuit 109
The vector esl is convolved with the impulse response of the filter shown below to calculate the vector escl. H (Z) = 1 / (1-εZ ^−L ) (5) where ε = 1.0 and L is the pitch lag obtained from the index of the adaptive excitation code. A shift type adaptive codebook (see Japanese Patent Laid-Open No. 01-54497)
Then, the index of the adaptive excitation code and the pitch lag are in one-to-one correspondence with each other as follows, for example. Adaptive excitation code index 0 1 2 ・ ・ ・ i ・ ・ ・ ↓ ↓ ↓ ↓ ↓ Pitch lag 20 21 22 ... 20 + i ...

[0020]

Since the statistical excitation vector is converted for each coding time, an excitation vector that is well suited to the input speech vector can be obtained even with a small number of vectors, and excellent synthesized speech quality can be obtained. This is effective for lowering the bit rate.

[Brief description of drawings]

FIG. 1 is a block diagram of a code-excited linear predictive coding apparatus according to the present invention.

FIG. 2 is a block diagram of a code-excited linear predictive decoding apparatus corresponding to FIG.

[Explanation of symbols]

 102 vocal tract analysis circuit 103 LPC quantizer 104 LPC inverse quantizer 105 synthesis filter 106 multiplexing circuit 107 adaptive excitation codebook 108 statistical excitation codebook 109 code vector conversion circuit 110 VQ gain codebook 111 perceptual filter 112 perceptual error calculation Circuit 113 Multiplier 114 Multiplier 115 Adder C Total code eai Adaptive excitation candidate vector esl Statistical excitation candidate vector Ia Adaptive excitation code index Ic LPC code Ig Gain code index Is Statistical excitation code index S Input speech vector Sw Synthetic speech vector βk Adaptive Excitation candidate gain γk Statistical excitation candidate gain

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

Claims (2)

[Claims] 1. A means for obtaining a vocal tract parameter by linear predictive analysis of an input speech vector, and a voice for quantizing the vocal tract parameter and outputting a quantized vocal tract parameter and a vocal tract parameter code corresponding thereto. A road parameter quantizer, an adaptive excitation codebook that stores an adaptive excitation vector that represents the optimal excitation vector of past input speech vectors, and a statistical excitation vector that is a predetermined excitation vector. A statistical excitation codebook; and a means for creating an excitation vector by multiplying each of the adaptive excitation vector and the statistical excitation vector by an excitation gain and adding the excitation gain, and the quantized vocal tract parameter and the excitation vector Create a synthesized speech vector based on the above, and evaluate the error between the input speech vector and the synthesized speech vector. By determining the index of both excitation codebooks corresponding to the excitation vector optimum to the input speech vector at the current time, the code excitation linear prediction code that outputs the vocal tract parameter and both indexes as a total code In the digitizing device, a specific impulse response is created based on the vocal tract parameter, and, using the impulse response, the statistical excitation vector is subjected to convolution processing, and then the adaptive excitation vector is added to the statistical excitation vector. A code excitation linear predictive coding method, comprising means for creating an excitation vector, wherein the impulse response corresponds to an impulse response of a transfer function expressed using the vocal tract parameter. 2. A means for obtaining a vocal tract parameter by linear predictive analysis of an input speech vector, and a voice which quantizes the vocal tract parameter and outputs a quantized vocal tract parameter and a vocal tract parameter code corresponding thereto. A road parameter quantizer, an adaptive excitation codebook that stores an adaptive excitation vector that represents an optimal excitation vector of past input speech vectors, and a statistical excitation vector that is a predetermined excitation vector. A statistical excitation codebook; and a means for creating an excitation vector by multiplying each of the adaptive excitation vector and the statistical excitation vector by an excitation gain and adding the excitation gain, and the quantized vocal tract parameter and the excitation vector Create a synthesized speech vector based on the above, and evaluate the error between the input speech vector and the synthesized speech vector. By determining the index of both excitation codebooks corresponding to the excitation vector optimum for the input speech vector at the current time, the code excitation linear prediction code that outputs the vocal tract parameter and both indexes as a total code In the digitizing device, a specific impulse response is created based on the vocal tract parameter, and, using the impulse response, the statistical excitation vector is subjected to convolution processing, and then added to the adaptive excitation vector. The impulse response is a transfer function H (Z) represented by the following equation.
A code-excited linear predictive coding device characterized by the impulse response of. H (Z) = 1 / (1-εZ ^−L ) where ε is a constant 0 <ε ≦ 1, and L is a pitch lag calculated from the index of the adaptive excitation vector.