JP3292227B2 - Code-excited linear predictive speech coding method and decoding method thereof - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention weights and synthesizes a vector that repeats a past excitation vector selected from an adaptive codebook at a pitch cycle and a noise vector selected from a noise codebook, and performs linear prediction synthesis using the synthesized vector. A code for encoding the input speech signal by selecting each vector from the adaptive codebook and the noise codebook so that the filter is driven to perform speech synthesis, and the distortion of the synthesized speech signal with respect to the input speech signal is minimized. The present invention relates to an excitation linear predictive speech coding method and a decoding method for decoding a decoded output by the coding method.

[0002]

2. Description of the Related Art In digital mobile communications and voice storage services, various high-efficiency voice coding methods are used in order to efficiently use radio waves and storage media. For example, code excitation linear prediction coding (Code Excited L)
inner Prediction, CELP), Vector Sum Excited Linear Prediction Coding (Vector Sum E)
xcited Linear Prediction,
VSELP) is known. For each technology, see M.E. R. Schroeder
and B. S. Atal: "Code-Excit
ed Linear Prediction (CEL
P): High-quality Speech at
Very Low Bit Rates ”, Pro
c. ICASSP-85, 25.1.1, pp. 937
-940, (1985); A. Gers
on and M. A. Jasiuk: "Vector
SumExcited Linear Predic
Tion (VSELP) Speech Coding
at 8 kbps ", Proc. ICASP-90,
S9.3, pp. 461-464 (1990).

[0003] In these, one frame is composed of about 5 ms to 50 ms, and one adaptive code vector in a pitch adaptive codebook composed of past excitation signals and a noise code composed of a fixed noise or a pulse train stored in advance. The weighted sum with the noise code vector of the book is used as the excitation signal. To minimize the auditory weighted waveform distortion between the input signal and the synthesized waveform obtained by passing the excitation signal through the linear prediction synthesis filter,
Determine the adaptive code, noise code, and gain code.

FIG. 5 shows a basic configuration of a CELP coding method as a conventional method. First, the input audio signal from the input terminal 1 is inputted as a digital signal, the input audio signal in the linear prediction analyzer 2 are linear prediction analysis of the speech is done, be further quantized, the coded output as prediction coefficients A It is set as a part and set as a coefficient of the synthesis filter 3. The adaptive codebook 4 includes the immediately preceding synthesis filter 3
The excitation source vector used as a driving input to the adaptive codebook 4 is stored, the vector in the adaptive codebook 4 is cut out, and the pitch corresponding to the code L of the adaptive codebook 4 is minimized so that the synthesized waveform distortion of the synthesis filter 3 is minimized. An adaptive code vector is repeated at a period T (L). That is, as shown in FIG. 6A, the past excitation signal Sp stored in the adaptive codebook 4 is cut out by the pitch period T (L), and this is repeated, that is, is cycled to obtain an adaptive code vector Va. In the adaptive code vector Va of various pitch periods T (L), the one that minimizes the distortion of the synthesized speech is selected.

[0005] A noise code vector composed of a plurality of noises or pulse trains is stored in the noise codebook 5 in advance, and a vector having the minimum combined waveform distortion is selected from the noise code vectors together with the adaptive code vector. Lastly, the adaptive code vector selected by the multipliers 7 and 8 and the selected noise code vector are multiplied by gains and weighted, and the weighted vectors are added to the adder 9.
, And the gain signals G0 and G1 of the multipliers 7 and 8 are also optimized to minimize the waveform distortion. The distortion with the auditory weight between the synthesized waveform from the synthesis filter 3 and the input speech waveform is calculated by the distortion calculator 1.
0, and the codebook search unit 11 controls the selection of each vector of the adaptive codebook 4 and the noise codebook 5 and the gain applied to the multipliers 7 and 8 so that this distortion is minimized. , The selection code (pitch code) L of the adaptive codebook 4, the selection code C of the noise codebook 5, and the gain code G
0, G1 and the quantized prediction coefficient A are coded output of input speech. This encoding is performed on a frame basis (5 to 5).
(About 50 milliseconds).

The noise codebook 5 further encodes the residual waveform of the excitation signal that could not be represented by the adaptive code vector. The configuration of the random codebook 5 initially includes 8
In encoding at about kbit / s, a random time-series vector such as Gaussian noise is generated and used. However, to reduce the bit rate to 4 kbit / s or less, it is necessary to reduce the bit allocation per vector or increase the number of dimensions of the vector. In this case, the residual signal expressed by the noise codebook is Since the signal is not necessarily random but contains a large number of pitch periodicity and pulses, diversification of the expression of the random codebook is becoming important.

As a technique for diversifying noise excitation sources, a pitch synchronous noise excitation source coding method (Pitch Syncor)
nous Innovation-CELP: PSI-
CELP). About this technology, Miki, Moriya,
Mano, Omuro: “CELP coding with pitch synchronous noise excitation source (PSI-CELP)”, IEICE Transactions, A, Vol. , J77-A, No. 3, pp. 31
4-324 (March 1994). This uses a pitch period T (L) obtained from the adaptive codebook 4 as shown in FIG. 6B, and adds a fixed noise vector Vsc selected from the noise codebook 5 from the top to T (L) in the same manner as the adaptive code vector Va. L), and this is repeated, that is, periodicized and used as the noise coded vector Vs. If the pitch cannot be obtained by the adaptive codebook 4, the noise vector is not periodized.

[0008] In PSI-CELP, the noise codebook is obtained by learning from a large-scale database. About this technology, Moriya, Miki, Omuro, Mano:
“Exercise Codebook Learning in CELP Coding”, IEICE Transactions, A, Vol. , J77-A, No.
3, pp. 485-493 (March 1994). This is a method that uses a vector quantizer design method based on a generalized Lloyd algorithm and uses a measure that minimizes auditory weighted waveform distortion during actual encoding.

[0009] These noise codebooks are always fixed in the case of CELP, and PSI-CEL
In P, the codebook before pitch periodization is configured to always generate a noise vector from the same noise codebook regardless of the pitch period.

[0010]

When the adaptive code vector has pitch periodicity, large power residuals are distributed around the pitch peak position. However, in the related art, the random codebook itself is fixed, and even if a method of making the random code vector pitch-period in accordance with the pitch period of the adaptive code vector is included in the random code vector. If the pitch periodicity is different from the pitch periodicity of the adaptive code vector, the coding efficiency cannot be sufficiently improved.

An object of the present invention is to provide a noise codebook with an adaptation function that includes a large number of noise code vectors suitable for each frame without increasing new transmission information, thereby diversifying excitation sources and providing high-quality speech. To provide a low bit rate audio encoding method for reproducing audio and a decoding method thereof.

[0012]

According to the encoding method of the present invention, a plurality of noise codebooks corresponding to a pitch period of an adaptive code vector are prepared, and a noise code corresponding to a pitch period selected by the adaptive codebook is prepared. A codebook is selected, and a random code vector is selected using the selected random codebook. The decoding method of the present invention prepares a plurality of noise codebooks in the same manner as the encoding method of the present invention, selects a noise codebook corresponding to the pitch period selected by the adaptive codebook, and selects the selected noise codebook. From the random code vector.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows an embodiment of the proposed speech coding method, and the same reference numerals in FIG. 5 denote the same parts. 1, a plurality of random codebook 5 ₁ to 5 in comparison with FIG. 5
_N are provided, the first noise codebook 5 ₁ to 5 _N switching unit 15 by the pitch period codes L from the adaptive codebook 4 is controlled
The difference is that one is selected. Noise codebook 5 ₁ to 5 _N is
For example, it is created by vector quantizer design based on the data for each range of L shown in FIG. 2A. At this time, L is T
It is assumed that numbers are assigned in ascending order of (L). Thus, the noise codebook 5 ₁ to 5 _N can be stored the random code vector corresponding to the pitch period. For example, a large number of learning voices are passed through an inverse filter to generate an adaptive codebook 4
Are collected, and those in the residual signal corresponding to the learning speech whose pitch period T (L) is in the range of T (l ₀ ) and T (l ₁ ) are collected, and are collected as noise codes. of creating a plurality of representative noise code vector as the learned data, which was used as a random codebook 5 _1, in the residual signal, a pitch period T (L) is T (l ₁₎ and T (l ₂₎ and the collect those corresponding range of training speech, which creates a plurality of representative noise code vector as a noise code training data, which was used as a random codebook 5 _2, similarly create other noise codebook below.

In order to encode the input speech by this encoding device, linear prediction analysis of the input speech is performed in the same manner as in the prior art.
An quantized prediction coefficient A is obtained and used as a coefficient of the synthesis filter 3, and an adaptive code vector periodicized by the pitch period T (L) is obtained from the vector of the adaptive codebook 4 so as to minimize the synthesized waveform distortion. On the basis of the obtained vector and the code L corresponding to the obtained vector, the switch 15 uses the correspondence table of FIG.
Determining the number of codebook to be used from the noise codebook 5 ₁ to 5 _N, selects the corresponding codebook 5 _i. From the noise codebook 5 _i selected in this way, first, assuming that the gain codes G0 and G1 of the multipliers 7 and 8 can take optimal values, the noise code vector C with the minimum waveform distortion is determined. Thereafter, with respect to the determined adaptive code vector and noise code vector C, the gain codes G0 and G1 are quantized so as to minimize the waveform distortion. Others are the same as those shown in FIG.

FIG. 2B shows an embodiment of the proposed speech decoding method, in which a code coded by the coding method shown in FIG. 1 is decoded. 4,
Noise codebook 5 ₁ to 5 _N and same adaptive each codebook 2
4. The random codebooks 25 _{1 to} 25 _N are provided to decode the input codes A, L, C, G0, and G1. The input quantized prediction coefficient A is set as a coefficient of the synthesis filter 23, and the multipliers 27 and 28 output gain codes G0 and G1 respectively.
Is set. As in the case of FIG. 1, the excitation source vector used as an input to the immediately preceding past synthesis filter 23 is stored in the adaptive codebook 24, and the pitch period determined by the input code L from the vector in the adaptive codebook 24 is stored. An adaptive code vector periodicized by T (L) is generated and output to the multiplier 27. Further, based on the value of the input code L, it is controlled by the switch 26 to control the noise codebooks 25 _{1 to} 25 _N
, A noise codebook to be used is selected, and a noise code vector of the input code C is selected from the selected noise codebook and output to the multiplier 28. The outputs of the multipliers 27 and 28 are added by an adder 29 and supplied as a drive signal to a synthesis filter 23. The output synthesized voice of the synthesis filter 23 is subjected to formant emphasis and pitch emphasis by a post filter 31 and the final voice The output is obtained.

FIG. 3 shows an embodiment of the speech encoding method according to claim 1 , in which parts corresponding to those in FIG. 1 are given the same numbers.
This device is the PSI-CELP described in the section of the prior art.
In the case where the present invention is applied, the adaptive codebook 4 has a fixed code in which, when there is no pitch periodicity, a noise vector fixed in advance is stored in addition to the code vector including the immediately preceding excitation source vector. A book 4b is provided,
A period control section 32 is provided to make the pitch cycle of the noise code vector output from the switch 61. Each codebook codebook 5 ₁ to 5 _N, which is made in the same manner as the noise codebook design of FIG. 1, therein is also random codebook corresponding to the case where there is no pitch periodicity to the adaptive code vector include.

The input speech is subjected to linear prediction analysis and quantization as in the case of FIG. 1, and the quantized prediction coefficient A becomes a coefficient of the synthesis filter 3. In the adaptive codebook 4, either the vector of the adaptive codebook 4a or the fixed noise vector in the fixed codebook 4b is selected so as to minimize the synthesized waveform distortion, and a code indicating the selected vector is selected. L
Includes a code indicating the pitch period and a code indicating the number of the fixed codebook 4b.

The switch 61 determines the number of the noise codebook to be used from the correspondence table shown in FIG. 2A based on the value of the code L, and selects the noise codebook. For example, where the noise codebook 5 ₁ ~5 _N-1 is the codebook when the sign L is a pitch periodicity, the noise codebook 5 _N shall correspond to the code L no pitch periodicity . When the code L is a code having a pitch periodicity with respect to the storage vector of the noise codebook selected in this way, the pitch noise control unit 32 converts the input noise vector as shown in FIG. By repeating with a pitch period T (L), a periodic noise vector is generated. By this operation, as a noise vector, the storage vector of the noise codebook having the pitch periodicity corresponding to the code L is periodicized at the corresponding pitch period, thereby increasing the number of vector candidates likely to reduce the quantization distortion. Can be generated. If the code L has no pitch periodicity, the noise vector is not periodic. That codebook 5 _N is selected, and supplies a random code vector selected from the noise code book 5 _N is to the multiplier 8 without periodic. Others are the same as those in FIG.

FIG. 4 shows an embodiment of the speech decoding method according to claim 2 , in which the same reference numerals are given to the parts corresponding to those in FIG. 2B. This decoding method is for decoding a code by the encoding method shown in FIG. Therefore, the adaptive codebook 4 is composed of the same adaptive codebook 24a and fixed codebook 24b as the adaptive codebook 4a and the fixed codebook 4b, respectively, and the pitch period control unit 33 is provided on the output side of the switch 26.

The input quantized prediction coefficient A is set as a coefficient of the synthesis filter 23. In the adaptive codebook 24, based on the input code L, the adaptive code vector from the adaptive codebook 24a or the adaptive codebook 24b from the fixed codebook 24b. One of the fixed noise vectors is selected, and the switch 26 determines the number of the noise codebook to be used from the correspondence table shown in FIG. Select With respect to the storage vector of the noise codebook thus selected, in the case where the code L is a code having a pitch periodicity, the selected noise vector is stored in the pitch period control unit 33 as shown in FIG. 6B. , And a pitch cycle to generate a periodic noise vector. If the code L has no pitch periodicity, the noise vector is not periodic. Then, the adaptive code vector and the noise code vector are multiplied by gains corresponding to the gain codes G0 and G1 in the multipliers 27 and 28, and are passed through the synthesis filter 23. The output of the synthesis filter 23 is subjected to formant emphasis and pitch emphasis by the post filter 31 to obtain a final audio output.

[0021]

As described above, the present invention comprises a plurality of random codebooks and switches between random codebooks depending on the pitch period obtained by the adaptive codebook or the presence or absence of pitch periodicity. Can be selected. Furthermore, it is not necessary to transmit new information because the pitch period selected by the adaptive codebook is used for switching the noise codebook, and without increasing the transmission information, the noise codebook is adapted to increase the coding efficiency, High quality speech coding is possible even at a low bit rate.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of a proposed speech encoding device.

[2] A is a diagram showing the correspondence between the noise codebook for switching the sign L from the adaptive codebook according to an embodiment of the invention, B is Hisage
It is a block diagram which shows the example of the devised audio | voice decoding apparatus.

FIG. 3 is a block diagram showing an example of a speech encoding apparatus to which the embodiment of the first invention is applied.

FIG. 4 is a block diagram showing an example of a speech decoding apparatus to which the embodiment of the invention of claim 2 is applied.

FIG. 5 is a block diagram showing a coding apparatus using a conventional code excitation linear prediction speech coding method.

FIG. 6A is a diagram illustrating a state in which an adaptive code vector is created from a vector stored in an adaptive codebook, and FIG. 6B is a diagram illustrating a state in which a random code vector is periodicized at a pitch.

Claims (2)

(57) [Claims] 1. A linear predictive synthesis filter is driven by a weighted sum of a vector selected from an adaptive codebook that repeats a past excitation vector at a pitch cycle and a vector selected from a noise codebook stored in advance. In the code-excitation linear predictive speech coding method for synthesizing a speech signal, selecting a vector from each of the codebooks so as to minimize distortion of the synthesized speech signal with respect to the input speech signal, and encoding the input speech signal. Removes the above adaptive codebook vector components from the training speech
Learned with signals within a predetermined pitch period in the signal
A plurality of random codebooks storing the random codebooks stored therein , and selecting one random codebook from the plurality of random codebooks in accordance with the pitch period of the vector selected in the adaptive codebook; and selects the vector from the noise codebook, said at a pitch period vector selected in the adaptive codebook
Periodize the vector selected from the random codebook and calculate the weight
Adds the sum, and the vector selected in the above adaptive codebook has no pitch period
If this is the case, select the multiple random codebooks
The selected vector from the selected random codebook
A code excitation linear predictive speech coding method characterized by not performing the above-mentioned periodization. 2. A vector is selected from an adaptive codebook that repeats a past excitation vector at a pitch cycle in accordance with an input code and a noise codebook that has been stored in advance. In a code excitation linear predictive speech decoding method for decoding a speech signal by driving a predictive synthesis filter, a vector component of the adaptive codebook is removed from a learning speech.
Learned with signals within a predetermined pitch period in the signal
Multiple noise codebooks containing random code vectors
A plurality of random codebooks are selected from the plurality of random codebooks according to the pitch period selected by the adaptive codebook , and the vector is selected from the selected random codebooks. Above with the pitch period of the selected vector
Periodize the vector selected from the random codebook and calculate the weight
Adds the sum, and the vector selected in the above adaptive codebook has no pitch period
If this is the case, select the multiple random codebooks
The selected vector from the selected random codebook
A code-excited linear predictive speech decoding method characterized by not performing the above-mentioned periodization.