JP3050966B2 - Code excitation linear predictive encoder - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses an adaptive excitation codebook, a statistical excitation codebook, and an excitation gain codebook. The present invention relates to a prediction encoding device.

[0002]

2. Description of the Related Art A code excitation linear predictive coding apparatus including an adaptive excitation codebook storing an optimal excitation vector of a past original speech vector and a statistical excitation codebook storing a predetermined excitation vector, For example, it is disclosed in the following document. Reference: N. S. Jayant and J.M. H. Ch
en, “SpeechCoding with Tim
e-Varying Bit Allocations
to Excitation and LPC Par
parameters, "Proc. ICASPS, pp65
-68, 1989. In this document, we consider that the residual signal for speech synthesis is composed of a periodic signal and a noisy signal,
A plurality of adaptive excitations read out from the statistical excitation codebook corresponding to the noisy signal, the adaptive excitation codebook corresponding to the periodicity signal, and the optimal excitation vector for the original speech vector are selected and determined. A plurality of excitation vectors are created based on the vector and the plurality of statistical excitation vectors. More specifically, the adaptive excitation code book stores a past excitation vector, that is, an optimal residual signal sequence to some extent, outputs a signal x samples in the past, and changes the value of x to generate a plurality of adaptive excitation vectors. (N
If there are x, n candidate vectors are obtained).

[0003]

However, in this case, independent excitation gain codebooks are prepared for the adaptive excitation vector and the statistical excitation vector. In other words, the excitation gain is scalar-quantized. Therefore, there is a problem that sufficient performance cannot be obtained when the number of quantization bits is reduced. Certainly, the statistical excitation vector is predetermined (the statistical excitation codebook is predetermined), whereas the power of the adaptive excitation vector varies at each time (the adaptive excitation codebook is (Updated at each encoding), so that there is almost no correlation between the powers. Therefore, generally speaking, the statistical excitation gain and the adaptive excitation gain are set independently. On the other hand, for any original speech vector, the adaptive excitation vector selected as optimal is eao, the adaptive excitation gain is βo, the statistical excitation vector is eso, and the statistical excitation gain is γ.
If o, βο * eao power (sum of squares of vector elements)
And the power of γo * eso have a considerable correlation (βο *
When the power of eao is high, the power of γo * eso is also high). Therefore, considering an appropriate gain g, γo * eso is calculated as γ
When expressed as o * g * eso and eao and g * eso are correlated, a correlation appears between the adaptive excitation gain βo and the statistical excitation gain γo. The present invention focuses on such correlation of excitation gains, and has as its object to provide a code excitation linear prediction encoding device that can obtain excellent synthesized speech quality even in low bit encoding. is there.

[0004]

The present invention accumulates optimum excitation vectors for a plurality of past original speech vectors and outputs them as adaptive excitation vectors under the control of selection control means. Has a codebook. Further, there is provided a statistical excitation codebook which accumulates a plurality of predetermined excitation vectors and outputs them as statistical excitation vectors under the control of the selection control means. Also, there is provided an excitation gain codebook which stores a plurality of predetermined excitation gains and outputs them as adaptive excitation gains and statistical excitation gains under the control of the selection control means. Further, there is provided an excitation vector synthesizing means for generating a plurality of excitation vectors based on the outputs of the adaptive excitation codebook, the statistical excitation codebook, and the excitation gain codebook under the control of the selection control means.
Further, a synthesized speech vector is created based on the vocal tract prediction parameters of the original speech vector and each excitation vector, and an error vector between the original speech vector and each synthesized speech vector is evaluated by predetermined evaluation means. There is selection control means for accumulating the minimum excitation vector as an optimal excitation vector in the adaptive excitation codebook. Then, the excitation vector synthesizing means multiplies the adaptive excitation vector by the adaptive excitation gain, and multiplies the statistical excitation vector by a value proportional to the square root of the power of the adaptive excitation vector and the statistical excitation gain. Each excitation vector is created by adding the adaptive excitation vector and the statistical excitation vector. In this specification, the square root of the power includes the sum of the magnitudes of the respective elements of the adaptive excitation vector, which is almost the same value.

[0005]

The excitation gain code book is obtained by accumulating the excitation gain by vector quantization, that is, the adaptive excitation gain β under the common gain code index Ig.
k and the statistical excitation gain γk. When each excitation vector e is created, the amplitude of the statistical excitation vector esl is changed by the correction gain gi shown in the equation (1). The vector e is represented by Expression (2). gi = α * RMS [eai] (1) where α is a constant, and RMS [eai] is the square root of the power of the adaptive excitation vector eai e = βk * eai + γk * gi * esl (2) Amplitude and statistics of the adaptive excitation vector eai By connecting the amplitude of the excitation vector esl with the correction gain gi,
Ideally speaking, by predicting the amplitude of the statistical excitation vector esl with the correction gain gi relating to the adaptive excitation vector eai, the optimal excitation vector (the excitation vector of the original speech vector) eopt between the two excitation gains βo and γo Shows a correlation. Therefore, by specifying the combination of the adaptive excitation gain βk and the statistical excitation gain γk with the common gain code index Ig, it is possible to perform low-bit encoding for the excitation gain rather than associating each independent gain code index. Since the value of the constant α in the equation (1) is related to the power of the statistical excitation vector esl itself, it should be experimentally determined. As a guide, the adaptive excitation gain dynamic range and the statistical excitation It is appropriate that the dynamic range of the gain is the same.

[0006]

FIG. 1 shows an embodiment of a code excitation linear prediction encoder according to the present invention. An original speech vector S that is grouped in frame units from the terminal 301 and input as a vector
Is first input to the vocal tract analysis circuit 313, and a vocal tract prediction parameter aj is calculated. The vocal tract analysis circuit 313 sends the vocal tract prediction parameter aj to the LPC quantizer 312. The LPC quantizer 312 calculates the vocal tract prediction parameters aj
, And sends the LPC code Ic to the LPC dequantizer 311 and the multiplexing circuit 317. The LPC inverse quantizer 311 converts the LPC code Ic into a vocal tract prediction parameter a
The signal is inversely converted to qj and sent to the synthesis filter 310. next,
The adaptive excitation codebook 302 contains the adaptive excitation vector eai
(I = 1 to n), the statistical excitation codebook 303 outputs the statistical excitation vector esl (l = 1 to m), and the VQ gain codebook 305 outputs the excitation gains βk and γk (k = 1 to p). The RMS calculation circuit 304 calculates the correction gain gi = α × R shown in Expression (1) based on the adaptive excitation vector eai.
Calculate and output MS [eai]. Adaptive excitation vector eai
Is multiplied by the adaptive excitation gain βk by the multiplier 308 to become a vector eaik, and the statistical excitation vector esl is
06 and the multiplier 307 multiply the correction gain gi and the statistical excitation gain γk to obtain a vector eslik.

[0007] The adder 309 calculates the excitation vector e by adding the component of the vector eaik and the vector eslik. Synthesis filter 310 calculates a synthesized speech vector S _W for excitation vector e, and sends to the subtractor 314. The subtractor 314 subtracts the components of the two vectors _SW and S from each other, and sends the error vector er to the perceptual filter 315. The perceptual filter 315 sends the output vector ew corresponding to the error vector er to the perceptual error calculation circuit 316. The perceptual error calculation circuit 316 calculates a root-mean-square of each component of the vector ew, and i, l, k that minimizes this value.
Index I of each codebook
a, Is, and Ig as the adaptive excitation codebook 302;
It is sent to the statistical excitation codebook 303, VQ gain codebook 305, and multiplexing circuit 317. The adaptive excitation codebook 302 includes an optimal adaptive excitation code vector eao and a statistical excitation codebook 303 based on the index Ia.
Is an optimal statistical excitation code vector eso according to the index Is, and the VQ gain codebook 305 outputs optimal VQ gains βo and γo according to the index Ig. The RMS calculation circuit 304 calculates and outputs the gain go based on the vector eao. The optimal excitation vector eopt constituted by these is stored in the adaptive excitation codebook 302.
And the contents of the adaptive excitation codebook 302 are updated. The multiplexing circuit 317 includes Ic, Ia, Is, and Ig.
Is transmitted to the receiving side through the output terminal 318 as the total code C.

In FIG. 2, at each time, first, vocal tract parameters aj, Ic and aqj are calculated from the original speech vector S (step ST1). Next, in each combination of i, l, and k, a correction coefficient gi (eai and gi correspond one-to-one) corresponding to the adaptive excitation vector eai is calculated,
Further, an excitation vector e is calculated (step ST2), and various vectors S _w , er, ew such as a synthesized speech vector and a perceptual error are calculated (step ST3). Do about. (Step ST4). Thereafter, each index that minimizes the perceptual error is output (step ST5), an optimal excitation vector eopt is constructed (step ST6), and the adaptive excitation codebook is updated with the excitation vector eopt (step S5).
T1).

FIG. 3 is a distribution diagram showing the relationship between the adaptive excitation gain βk and the statistical excitation gain γk for a group of words encoded by the code-excitation linear predictive encoding apparatus of FIG. As apparent from FIG. 3, the adaptive excitation gain β
Since there is a strong correlation between k and the statistical excitation gain γk,
By sampling representative points of the distribution and storing them in the VQ gain codebook of FIG.
g can be reduced in bit.

FIG. 4 shows an embodiment of an excitation linear prediction encoding apparatus according to the present invention. The code C input from the input terminal 401 is converted by the demultiplexing circuit 413 into the LPC code Ic,
These are separated into an adaptive excitation code index Ia, a statistical excitation code index Is, and a gain code index Ig, and sent to the LPC dequantizer 411, the adaptive excitation codebook 402, the statistical excitation codebook 403, and the VQ gain codebook 405, respectively. LPC dequantizer 41
1 converts the LPC code Ic into a vocal tract prediction parameter aj and sends it to the synthesis filter 410. The adaptive excitation codebook 402 stores the adaptive excitation code vector ea corresponding to the index Ia, the statistical excitation codebook 403 stores the statistical excitation vector es corresponding to the index Is, and the VQ gain codebook 405 stores the excitation gains β and γ corresponding to the index Ig. Output each. RMS calculation circuit 40
4 calculates and outputs the gain g based on the vector ea. An excitation vector e is constituted by ea, es, β, γ, and g. The synthesis filter 410 calculates a synthesized speech vector S for the vector e and outputs it from the output terminal 412. The contents of the adaptive excitation codebook 402 are updated by the vector e.

[0011]

As described above, since the gains of the two codebooks are vector-quantized, the quantization characteristics are better than those of the scalar quantization. Further, the use of the RMS calculation circuit allows a higher correlation between the gains of the two codebooks. Appear, the effect of vector quantization is further enhanced. These are effective for lowering the bit rate.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention.

FIG. 2 is a flowchart illustrating the operation of FIG.

FIG. 3 is a diagram illustrating a relationship between an adaptive excitation gain and a statistical excitation gain in the configuration of FIG. 1;

FIG. 4 is a block diagram of a decoding device according to the present invention.

[Explanation of symbols]

 302 Adaptive excitation codebook 303 Statistical excitation codebook 304 RMS calculation circuit 305 VQ gain codebook 306 Multiplier 307 Multiplier 308 Multiplier 309 Adder 310 Synthesis filter 311 LPC inverse quantizer 312 LPC quantizer 313 Vocal tract analysis circuit 314 Subtractor 315 Perceptual filter 316 Perceptual error calculation circuit

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-1-54497 (JP, A) JP-A-4-270400 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 H03M 7/30 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. An adaptive excitation codebook which accumulates optimal excitation vectors for a plurality of past original speech vectors and outputs them as an adaptive excitation vector under the control of selection control means. Accumulate the specified excitation vector,
They are controlled by the selection control means and output a statistical excitation codebook as a statistical excitation vector, and a plurality of, accumulate a plurality of predetermined excitation gain, which are controlled by the selection control means An excitation gain codebook to be output as an adaptive excitation gain and a statistical excitation gain, and controlled by the selection control means, based on the outputs of the adaptive excitation codebook, the statistical excitation codebook, and the excitation gain codebook, Excitation vector synthesizing means for generating a plurality of excitation vectors; generating a synthesized speech vector based on the vocal tract prediction parameters of the original speech vector and each of the excitation vectors; and calculating an error between the original speech vector and each of the synthesized speech vectors. The vector is evaluated by a predetermined evaluation means, and the excitation vector having the minimum evaluation value is determined as an optimal value. The selection control means accumulating in the adaptive excitation codebook as an excitation vector, corresponding to the optimal excitation vector, the index of the adaptive excitation vector, the index of the statistical excitation vector, and the index of the excitation gain, and In the code excitation linear prediction encoding device, wherein the information relating to the vocal tract prediction parameters is encoding information, the excitation vector synthesis unit multiplies the adaptive excitation vector by the adaptive excitation gain, and the statistical excitation vector Multiplying a value proportional to the square root of the power of the adaptive excitation vector and the statistical excitation gain, and adding the multiplied adaptive excitation vector and the statistical excitation vector to generate each of the excitation vectors. Characteristic code-excited linear prediction encoding device.