JP3068688B2 - Code-excited linear prediction coding method - Google Patents

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 prediction encoding method.

[0002]

2. Description of the Related Art The following document discloses a code excitation linear prediction including an adaptive excitation codebook storing an optimal excitation vector for a past input speech vector and a statistical excitation codebook storing a predetermined excitation vector. An encoding technique is disclosed. Reference: N. S. Jayant and J.M. H. Ch
en, “SpeechCoding with Tim
e-Varying Bit Allocations
to Excitation and LPC Pa
parameters ", Proc. ICASP, pp6
5-68, 1989. The present inventors have previously described a code-excitation linear predictive encoding apparatus having a VQ gain codebook in which the gains of two excitation codebooks are vector-quantized, as disclosed in Japanese Patent Application No.
56067 was filed as an invention. In this case, the vector quantization of the gain is effectively performed by giving a high correlation between the gains of the two excitation codebooks using an RMS calculation circuit.

[0003]

However, since the RMS calculation circuit predicts the gain of the statistical excitation code based on the adaptive excitation vector, there is no need to transmit power information to the decoder side. If the index of the adaptive excitation codebook is incorrect due to a transmission path error, there is a problem that the synthesized voice quality is greatly deteriorated and is vulnerable to a transmission path error. Accordingly, the present invention provides a code-excited linear predictive encoding method that is resistant to transmission path errors without significantly degrading synthesized speech quality even if an error occurs in the transmission path between the encoder and the decoder. Aim.

[0004]

According to the present invention, a first excitation vector is calculated by multiplying an adaptive excitation vector by a gain of a VQ (vector quantization) codebook. Further, a correction gain is calculated by performing a leak process on the gain calculated from the average power value of the adaptive excitation vector. Then, the second excitation vector is calculated by multiplying the statistical excitation vector by the correction gain and the gain of the VQ codebook. Further, a first excitation vector and the second excitation vector are added to calculate an excitation vector candidate, and a difference between a synthesized speech vector obtained using the excitation vector candidate and an input speech vector is calculated. An excitation vector candidate that minimizes an error obtained by hearing weighting is determined as an optimal excitation vector at the current time.

[0005]

The VQ gain codebook is obtained by accumulating an excitation gain by vector quantization and outputting an adaptive excitation gain βk and a statistical excitation gate γk under a common gain code index Ig. Further, the RMS value gi represented by the equation (1) is calculated, and the RMS value gi is calculated.
, A leak process shown in Expression (2) is performed to create a correction gain Gi. Then, when creating each excitation candidate vector e, the amplitude of the statistical excitation vector esl is changed by the correction gain Gi and the statistical excitation gain γk. 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. The constant α can be determined by simulation.

Gi = (1−D) * G (n) + D * gi (2) where D is a constant and G (n) is a value calculated based on the optimal adaptive excitation code vector at the previous frame time. This is the correction gain for the frame time. In an embodiment to be described later, D = ^2-7 . Thus, conceptually, the amplitude of the statistical excitation vector esl is predicted by the RMS value gi related to the adaptive excitation vector eai, and the optimal excitation vector (the excitation vector of the input speech vector)
A correlation appears between the two excitation gains βo and γo in eopt. Further, by performing the leak processing to generate the correction gain Gi, the transmission path error is dispersed. In this specification, as shown in Expression (2),
Corrected gain G (n) of previous frame and RMS of current frame
Weighted addition with the value gi is called leak processing. Also, the constant D
Is determined by estimating the transmission path error. As the error becomes smaller, the influence of the error continues for a longer time, but the influence of the error at the current time becomes smaller.

[0007]

FIG. 1 is a block diagram of a code excitation linear predictive encoder to which the present invention is applied. In FIG. 1, an input speech vector S that is grouped in frame units from a terminal 101 and input as a vector is first input to a vocal tract analysis circuit 113, and a vocal tract prediction parameter aj is calculated. The vocal tract analysis circuit 113 sends the vocal tract prediction parameter aj to the LPC quantizer 112. The LPC quantizer 112 quantizes the vocal tract prediction parameter aj and converts the code Ic (LPC code) into the LPC inverse quantizer 11.
1. Send to multiplexing circuit 117. LPC inverse quantizer 1
11 is to convert the LPC code Ic to the vocal tract prediction parameter aqj
, And sends the result to the synthesis filter 110.

Next, the adaptive excitation codebook 102 is an adaptive excitation vector eai (i = 1 to n), the statistical excitation codebook 103 is a statistical excitation vector esl (l = 1 to m),
The VQ gain codebook 105 outputs adaptive and statistical excitation gains βk and γk (k = 1 to p), respectively. The RMS calculation circuit 104 calculates RM based on the adaptive excitation vector eai.
The S value gi is calculated by the above equation (1), and the leak circuit 1
18 is output. The leak circuit 118 performs a leak process shown in the above equation (2) based on the RMS value gi to calculate and output a corrected gain Gi. Here, D = 2
^{As -7} , the correction gain Gi changes gradually.

The adaptive excitation vector eai is multiplied by an adaptive excitation gain βk by a multiplier 108 to obtain a first excitation vector eai.
aik, and the statistical excitation vector esl is
Is multiplied by the statistical excitation gain γk by the multiplier 107 to become the second excitation vector eslik. The adder 109 calculates the first excitation vector eaik and the second
The excitation vector eslik is added in component units to calculate an excitation vector candidate e. The synthesis filter 110 calculates a synthesized speech vector Sw for the excitation vector candidate e, and sends it to the subtractor 114. The subtractor 114 subtracts the synthesized speech vector Sw from the input speech S in component units, and sends the difference vector er to the perceptual filter 115.
The perceptual filter 115 sends an error vector ew for the difference vector er to the perceptual error calculation circuit 116.

The perceptual error calculation circuit 116 calculates the root mean square of each component of the error vector ew, and determines the excitation vector candidate (ie, the combination of i, l, k) having the minimum value, The optimal excitation vector of the input speech vector is determined. Then, the indexes Ia, Is, and Ig of the respective codebooks at that time are sent to the adaptive excitation codebook 102, the statistical excitation codebook 103, the VQ gain codebook 105, and the multiplexing circuit 117.

The adaptive excitation codebook 102 uses the index Ia to determine the optimal adaptive excitation vector eao, the statistical excitation codebook 103 uses the index Is to determine the optimal statistical excitation vector eso, and the VQ gain codebook 1
05 is the optimum VQ gain βo according to the index Ig,
γo are output. The RMS calculation circuit 104 calculates the optimum RMS value go based on the optimum adaptive excitation vector eao, and the leak circuit 118 calculates and outputs the optimum correction gain Go based on the optimum RMS value go. The optimal excitation vector eopt constituted by these is input to the adaptive excitation codebook 102, and the adaptive excitation codebook 1
02 is updated. The multiplexing circuit 117 transmits the indexes Ia, Is, Ig, and the LPC code Ic as a total code C to the receiving side via the output terminal 119.

FIG. 2 is a block diagram showing a code excitation linear prediction decoder corresponding to the encoder shown in FIG. In FIG. 2, a code C input from an input terminal 201 is demultiplexed by an LPC code Ic, an adaptive excitation code index Ia, a statistical excitation code index Is, V
Q gain code index Ig
C inverse quantizer 211, adaptive excitation codebook 202, statistical excitation codebook 203, VQ gain codebook 2
05.

The LPC inverse quantizer 211 converts the code Ic into a vocal tract prediction parameter aj,
To send to. The adaptive excitation codebook 202 is an adaptive excitation vector ea corresponding to the index Ia, the statistical excitation codebook 203 is a statistical excitation vector es corresponding to the index Is, and the VQ gain codebook 205 is an excitation gain β, γ corresponding to the index Ig. Are respectively output.

The RMS calculation circuit 204 calculates an RMS value g based on the vector ea.
The correction gain G is calculated and output based on the S value g. ea,
An excitation vector e is composed of es, β, γ, and G.
The synthesis filter 210 calculates a synthesized speech vector S for the vector e and outputs it from the output terminal 212. The contents of the adaptive excitation codebook 202 are updated by the vector e.

[0015]

By using a circuit that leaks the output of the RMS calculation circuit, power information is not transmitted to the decoder side, and an error is transmitted to the transmitter between the encoder and the decoder. Does not significantly degrade the synthesized speech quality.

[Brief description of the drawings]

FIG. 1 is a block diagram of a code excitation linear prediction encoding apparatus to which the present invention is applied.

FIG. 2 is a block diagram of a code excitation linear prediction decoding device corresponding to FIG.

[Explanation of symbols]

 Reference Signs List 102 adaptive excitation codebook 103 statistical excitation codebook 104 RMS calculation circuit 105 VQ gain codebook 106 multiplier 107 multiplier 108 multiplier 109 adder 110 synthesis filter 111 LPC inverse quantizer 112 LPC quantizer 113 vocal tract analysis circuit 114 Subtractor 115 Perception filter 116 Perception error calculation circuit C Total code e Excitation vector candidate eai Adaptive excitation vector candidate esl Statistical excitation vector candidate Gi Modified gain Ia Adaptive excitation code index Ic Vocal tract parameter code Ig Gain code index Is Statistical excitation code index S Input speech vector Sw Synthesized speech vector βk Adaptive excitation gain γk Statistical excitation gain

────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Yoshihiro Ariyama 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (56) References JP-A-4-270400 (JP, A) Kaihei 4-73700 (JP, A) Patent 3050966 (JP, B2) Proceedings of IE 1989 International Conference on Acoustics, Speech and Signal Processing, Vol. 1, "S2.7 Speech Coding with Time-Variing Bit Allocations to Excitatio and LPC Parameters" p. 65-68 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 11/00-13/08 G10L 19/00-21/06 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. An adaptive excitation codebook storing past optimal excitation vectors as adaptive excitation vectors, a statistical excitation codebook storing predetermined excitation vectors as statistical excitation vectors, and two And a VQ codebook storing the gain of the excitation codebook by vector quantization and updating the vocal tract prediction parameters according to the input speech vector in the code excitation linear prediction encoding method. A first excitation vector is calculated by multiplying the adaptive excitation vector of the codebook by the gain of the VQ codebook, and a leak process is performed on a gain calculated from an average power value of the adaptive excitation vector of the adaptive excitation codebook. To calculate the corrected gain, and to the statistical excitation vector of the statistical excitation codebook,
Multiplying the corrected gain by the gain of the VQ codebook to calculate a second excitation vector; adding the first excitation vector and the second excitation vector to calculate an excitation vector candidate; Calculate the difference between the synthesized speech vector obtained using the candidate and the input speech vector, determine the excitation vector candidate that minimizes the error obtained by auditory weighting the difference as the optimal excitation vector at the current time, A code excitation linear prediction encoding method characterized by sending an index of an excitation vector of each codebook and a vocal tract prediction parameter code at that time to a decoder side.