JPH05108096A - Vector drive type speech encoding device - Google Patents

Abstract

PURPOSE:To improve the quality of a reproduced speech even when the bit rate is low. CONSTITUTION:A 1st sound source which uses the components of the pitch vector of an adaptive code book 12a and the components of the code vector of a noise code book 14a, a 2nd sound source which uses only the components of the code vector, and a 3rd sound source which uses only the components of the pitch vector are generated as driving sound sources so as to minimize the error between a weighted input signal and a reproduced signal. For the purpose, an adaptive bit assignment part 22a adaptively assigns bit information to the index of the pitch vector and a multiplier 16a, and the index of the code vector and a multiplier 16b to vary the search ranges of the indexes of the code books and the numbers of quantized bits of the gains of the multipliers. Then a driving sound source signal determination part 18 determines the best driving sound source and the gain of the driving sound source.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector drive type speech coding apparatus, and more particularly to an input speech of 4.8 kb.
The present invention relates to a vector-driven type audio encoding device that encodes at a low bit rate of about / s.

[0002]

2. Description of the Related Art As an effective method for encoding a voice signal with good reproducibility and a low bit rate, CELP (Code
A vector-driven encoder based on the Excited Linear Prediction) encoding method has been proposed. For details of this method, see ICASSP-85pp. 937-940
B. S. “Code Excit” by Atal et al.
ed Linear Prediction (CELP): High QualitySpee
ch at Very Low Bit Rates. ”, IEICE Technical Report CS89-6
4 "Spindle Vector Driven Speech Coding" by Taniguchi et al., Or "Adaptive Codebook for CELP Coding by Taniguchi et al. Improvements ”.

An example of a conventional speech coding apparatus based on the CELP coding method is shown in FIGS. 3 and 4. This conventional speech coding apparatus 1 includes a coding apparatus 1a shown in FIG. 3 and a coding apparatus 1 shown in FIG.
The decoding device 1b shown in FIG. The coding device 1a has an adaptive codebook 2a and a noise codebook 3a, and the decoding device 1b has an adaptive codebook 2b and a noise codebook 3b. The adaptive codebooks 2a and 2b respectively record the same voice pattern,
The noise codebooks 3a and 3b also record the same voice pattern.

In the encoding device 1a shown in FIG. 3, the pitch vector (P) extracted from the adaptive codebook 2a having a predetermined number of indexes and also the noise vector 3a having a predetermined number of indexes are extracted. A driving sound source signal (bP + gC) is generated by multiplying the generated code vector (C) by each gain (pitch: b, code: g) and adding them. In the encoding process, the power of the error signal (E) between the reproduction signal (bAP + gAC) obtained by passing the driving excitation signal through the weighting synthesis filter 4a and the weighted input signal (AX) is used as an evaluation function in the error power evaluation unit 5. , A codebook search is performed. The number of bits of the index for the codebook search at this time is predetermined according to the number of codebook indexes. The driving sound source signal that minimizes the power of the error signal is the optimum driving sound source signal (b ₀ P
₀ + g ₀ C ₀ ). At this time, the contents of the adaptive codebook 2a are delayed by feeding back the optimum driving sound source signal to the adaptive codebook 2 through the frame delay unit by delaying the input signal by one frame divided for example every 7.5 ms. Will be updated.

The decoding device 1b shown in FIG.
Pitch vector (P₀)of
Index and gain (b₀), Noise Codebook
Code vector (C₀) Index
And gain (g₀) And weighted composite fill
The filter coefficient of the filter 4 is sent through the transmission line 7. That
Then, in the decoding device 1b, the optimum driving sound source is
Signal (b₀P₀+ G₀C₀) And create a weighted composite
Playback signal (b₀AP₀+ G ₀A
C₀) Is output.

[0006]

In the prior art, it is assumed that the pitch vector extracted from the adaptive codebook 2a reproduces a pitch period, and a voiced sound source is assumed. Therefore, the effect is great for voiced sounds, but less effective for non-periodic sounds such as unvoiced sounds. On the other hand, the code vector extracted from the noise codebook 3a is based on the assumption of an aperiodic unvoiced sound source, and has little effect on a periodic voice such as a voiced sound. Therefore, when the transfer rate is about 4.8 kb / s, the reproduced sound quality is not sufficiently satisfactory.

Therefore, a main object of the present invention is to provide a vector drive type audio encoding device capable of improving reproduced sound quality even at a low bit rate of, for example, about 4.8 kb / s.

[0008]

According to the present invention, a first sound source that uses, as a driving sound source, a pitch vector component extracted from an adaptive codebook and a code vector component extracted from a noise codebook, A second sound source that uses only the code vector components of the noise codebook;
A driving sound source creating means for creating a third sound source that uses only the components of the pitch vector of the adaptive codebook, and an information amount according to the first, second and third sound sources, and an index of the pitch vector of the adaptive codebook and By adaptive bit allocation means for adaptively allocating to gain and code vector index and gain of the noise codebook, comparison means for comparing the reproduction signal obtained by passing the driving sound source through the weighting synthesis means, and the weighting input signal, and the comparison means It is a vector-driven speech encoding apparatus including a drive excitation determination unit that determines any one of the first, second, and third excitations according to the comparison result.

[0009]

The driving sound source creating means uses the sound source in the first state in which both the pitch vector component extracted from the adaptive codebook and the code vector component of the noise codebook are used as the driving sound source, and the noise is used as the driving sound source. A sound source in the second state that uses only the components of the code vector extracted from the codebook, and a third that uses only the components of the pitch vector extracted from the adaptive codebook as the driving sound source.
Create a sound source in the state of. Therefore, the adaptive bit allocation means allocates the bit information amount of the index and the quantization bit number of the gain given to each codebook so that the optimum driving sound source matches each state of the first, second and third sound sources. change. In the case of the first sound source, the adaptive bit allocation means allocates a predetermined bit information amount to the pitch vector index of the adaptive codebook and its gain and the code vector index of the noise codebook and its gain. In the case of the second sound source, since the pitch vector component of the adaptive codebook is not used as the driving sound source, the bit information amount assigned to the pitch vector and the gain in the first state is used as the code vector of the noise codebook. Assign only to the index and gain of. This increases the amount of bit information assigned to the index and gain of the code vector of the noise codebook, widens the search range of the index when extracting the second sound source, and reduces the number of quantization bits for quantizing the gain. It fluctuates and an optimum second driving sound source is created. In the case of the third sound source, since the code vector component of the noise codebook is not used as the driving sound source, the information amount assigned to it in the first state is used only as the index and gain of the pitch vector of the adaptive codebook. assign. This increases the amount of information assigned to the pitch vector index and gain of the adaptive codebook, widens the search range of the index when extracting the first sound source, and changes the number of quantization bits for quantizing the gain. Then, an optimal third sound source is created.
In this way, the optimum driving sound source can be obtained by changing the allocation method of the bit information amount and creating the driving sound source in the state where the error power between the reproduction signal and the weighted input signal is minimized.

[0010]

According to the present invention, an optimum driving sound source can be produced regardless of whether the sound source is a voiced sound source or an unvoiced sound source, and therefore reproduction at a low bit rate of, for example, about 4.8 kb / s. The sound quality is improved compared to the conventional one. The above-mentioned objects, other objects, features and advantages of the present invention will become more apparent from the following detailed description of the embodiments with reference to the drawings.

[0011]

1 and 2, a vector driven speech coder 10 of this embodiment includes a coder 10a and a decoder 10b. The encoding device 10a has an adaptive codebook 12a for recording a pitch vector and a noise codebook 14a for recording a code vector, and the pitch vector (P) extracted from the adaptive codebook 12a is input to a multiplier 20a to generate noise. The code vector (C) extracted from the codebook 14a is input to the multiplier 20b. The multiplier 16a multiplies the pitch vector (P) by the gain (b) given by the driving sound source signal determining unit 18, and the multiplier 16b multiplies the code vector (C) by the gain given by the driving sound source signal determining unit 18. (G) is multiplied. Then, the outputs from the multipliers 16a and 16b are added in the adder 20 to generate the driving sound source signal (bP + gC).

The adaptive codebook 12a, the noise codebook 14a, the multiplier 16a and the multiplier 16b are provided with a predetermined bit information amount from the adaptive bit allocation unit 22a. The adaptive bit allocation unit 22a uses both the pitch vector (P) component of the adaptive codebook and the code vector (C) component of the noise codebook as the driving sound source, and the code of the noise codebook as the driving sound source. The bit information allocation method is changed depending on the state b in which only the component of the vector (C) is used and the state c in which only the component of the pitch vector (P) of the adaptive codebook is used as the driving sound source. As a result, the search ranges of the indexes of the adaptive codebook 12a and the noise codebook 14a change, and the quantization bit numbers of the multipliers 16a and 16b also change.

The driving sound source signal output from the adder 20 is reproduced through the weighting synthesis filter 24a (bAP).
+ GAC) and is input to the comparator 26. Comparator 2
At 6, the reproduced signal is compared with the weighted input signal (AX), and the error is input to the drive sound source signal determination unit 18 as error power (E). The driving sound source signal determination unit 18 selects the above states a to c so that this error power is minimized, and selects one pitch vector and one code vector from the adaptive codebook 12a and the noise codebook 14a, respectively. Only one of them is selected, and the gain to be multiplied to each or either one is determined. Then, the driving sound source signal that minimizes the error power is determined as the optimum driving sound source signal, and this is fed back to the adaptive codebook 12a via the frame delay unit 28a, and the contents of the adaptive codebook 12a are updated.

In determining the driving sound source, the gain (b) for the pitch vector (P) component of the adaptive codebook 12a is not "0" and the gain (b) for the code vector (C) component of the noise codebook 14a ( If g) is also not "0", it is optimal to code in state a. In this case, the predetermined bit information amount is determined by the adaptive bit allocation unit 22a as the index and gain (b) of the pitch vector (P) of the adaptive codebook 12a and the code vector (C) of the noise codebook 14a.
Assigned to the index and gain (g).
Then, the pitch vector (P) extracted from the adaptive codebook 12a and the code vector (C) extracted from the noise codebook 14a are multiplied by respective gains in the multipliers 16a and 16b and added in the adder 20. A driving sound source signal (bP + gC) is generated.

In the encoding process, the drive signal is reproduced signal (bA) obtained through the weighting synthesis filter 24a.
P + gAC) and the weighted input signal (AX) are compared 2
Compare with 6. Then, the drive sound source signal determination unit 18 uses the power of the error signal (E) output as the comparison result as an evaluation function so that the power of the error signal (E) is minimized.
Adaptive codebook 12a and noise codebook 14a
To determine the optimum pitch vector (P ₀ ) and the optimum code vector (C ₀ ), and further to determine the optimum pitch vector gain (b ₀ ) and the optimum code vector gain (g ₀ ). .. Then, a driving sound source signal is created from these, and the optimum driving sound source signal A (b ₀ P ₀ +
g ₀ C ₀ ).

When determining the driving sound source, when the gain (b) of the pitch vector is "0" and the gain (g) of the code vector is not "0", it is optimal to code in the state b. .. In this case, the adaptive bit allocation unit 22a sets the bit information amount allocated to the index and the gain (b) of the pitch vector (P) of the adaptive codebook 12a in the state a to “0”, and the corresponding amount is set to the noise code. It is assigned to the index and gain (g) of the code vector (C) of the book 14a. This expands the search range of the noise codebook 14a. Then, the multiplier 16b multiplies the code vector (C) extracted from the noise codebook 14a by the gain (g) to generate the driving sound source signal (gC).

In the encoding process, a reproduction signal (gA) obtained by the weighting synthesis filter 24a is added to the driving sound source signal.
C) and the weighted input signal (AX) are compared by the comparator 26. Then, the driving sound source signal determination unit 18 uses the power of the error signal (E) output as the comparison result as an evaluation function.
So as to minimize the power of this error signal (E), the noise codebook 14a is searched to determine the optimum code vector (C ₁ ) and further the optimum code vector gain (g ₁ ) is determined. To do. Then, a driving sound source signal is created from these, and is determined as the optimum driving sound source signal B (g ₁ C ₁ ).

When determining the driving sound source, when the pitch vector gain (b) is not "0" and the code vector gain (g) is "0", it is optimal to code in the state c. is there. In this case, the adaptive bit allocation means 22 sets the amount of information allocated to the index and the gain (g) of the code vector (C) of the noise codebook 14a in the state a to "0", and the corresponding amount is adapted codebook 12a. Of the pitch vector (P) and the gain (b). This widens the search range of the adaptive codebook 12a. Then, the multiplier 16a multiplies the pitch vector (P) extracted from the adaptive codebook 12a by the gain (b) to generate the driving sound source signal (bP).

In the encoding process, a reproduction signal (bA) obtained by the weighting synthesis filter 24a is added to the drive excitation signal.
P) and the weighted input signal (AX) are compared by the comparator 26. Then, the driving sound source signal determination unit 18 uses the power of the error signal (E) output as the comparison result as an evaluation function.
So as to minimize the power of this error signal (E), the adaptive codebook 12a is searched to determine the optimum pitch vector (P ₂ ) and further the optimum pitch vector gain (b ₂ ) is determined. To do. Then, a driving sound source signal is created from these, and is determined as the optimum driving sound source signal C (b ₂ P ₂ ).

Among the optimum driving excitation signals A to C thus determined, the one having the smallest error signal power is finally sent to the decoding device 10b through the transmission line 30 to obtain the optimum driving excitation signal. Used as Y. When the optimum driving sound source signal A is used as the optimum driving sound source signal Y,
The optimum code vector (P ₀ ) is sent from the adaptive codebook 12a to the adaptive codebook 12b, and the optimum code vector (C ₀ ) is sent from the noise codebook 14a to the noise codebook 14b. The gain of the multiplier 16a from the multiplier 16c to the optimum pitch vector (P ₀₎ (b ₀₎ are sent, the gain of the optimal code vector (C ₀₎ to the multiplier 16d from multiplier 16b (g ₀₎ Will be sent. Further, the filter coefficients are sent from the weighting synthesis filter 24a to the weighting synthesis filter 24b. Then, the adaptive bit allocation unit 22b causes the multipliers 16c and 1
It is determined from the gain input from 6d that the state is a, and the adaptive codebook 12b and the noise codebook 1
Bit information is given so that P ₀ and C ₀ are extracted for 4b. _{Then, P 0, C 0, b} 0 and g ₀ optimal excitation signal _{Y (b 0 P 0 + g} 0 C 0) is created from, through the weighting synthesis filter 24b reproduced signal (A
Y) is obtained.

The optimum driving sound source signal B is the optimum driving sound source signal Y.
Noise codebook 14a?
Optimal code vector (C
₁) Is sent. In addition, the multipliers 16b to 16d
Code vector (C ₁) Gain (g₁) Sent
Be done. In addition, the weighting synthesis filter 24a
The filter coefficient is sent to the attached synthesis filter 24b. So
Then, the adaptive bit allocation unit 22b causes the multiplier 16c and
From the gain input from 16d, determine that it is in state b.
Turn off and C for noise codebook 14b₁Is extracted
To give the bit information. And C₁And g₁
From the optimum driving sound source signal Y (g₁C₁) Is created and weights
The reproduction signal (AY) is obtained through the attachment synthesis filter 24b.
Be done.

The optimum driving sound source signal C is the optimum driving sound source signal Y.
Noise codebook 12a?
Optimal pitch vector (P
₂) Is sent. In addition, the multipliers 16a to 16c
Optimum pitch vector (P ₂) Gain (b₂) Sent
Be done. In addition, the weighting synthesis filter 24a
The filter coefficient is sent to the attached synthesis filter 24b. So
Then, the adaptive bit allocation unit 22b causes the multiplier 16c and
From the gain input from 16d, determine that it is in state c.
No, P for noise codebook 12b₂Is extracted
To give the bit information. And P₂And b₂
From the optimum driving sound source signal Y (b₂P₂) Is created and weights
The reproduction signal (AY) is obtained through the attachment synthesis filter 24b.
Be done.

Also in the decoding device 10b, the optimum driving excitation signal Y is fed back to the adaptive codebook 12b via the frame delay unit 28b, and the recorded content of the adaptive codebook 12b becomes the same as the recorded content of the adaptive codebook 12a. Will be updated. Further, in the above-mentioned embodiment, the vector-driven type speech coding apparatus has been described, but it is also possible to perform coding and decoding by computer simulation using the same method. In this case, the operation of the driving sound source determination unit 18, the adaptive bit allocation unit 22a, and the adaptive bit allocation unit 22b is performed by software, and the operation of other components is performed by hardware or software. Therefore, the error signal used for determining the driving sound source signal is not limited to the power.
In this case, needless to say, it is necessary to perform analog / digital conversion on the weighted input signal before the encoding.

[Brief description of drawings]

FIG. 1 is a block diagram showing an encoding device of a vector-driven speech encoding device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a decoding device of a vector-driven voice encoding device according to an embodiment of the present invention.

FIG. 3 is a block diagram showing an encoding device of a conventional vector-driven speech encoding device.

FIG. 4 is a block diagram showing a decoding device of a conventional vector-driven voice encoding device.

[Explanation of symbols]

 10 ... Vector-driven speech coding device 10a ... Encoding device 10b ... Decoding device 12a, 12b ... Adaptive codebook 14a, 14b ... Noise codebook 16a, 16b, 16c, 16d ... Multiplier 18 ... Driving excitation signal determination unit 20 ... Adder 22a, 22b ... Adaptive bit allocation part 24a, 24b ... Weighting synthesis filter 26 ... Comparator

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shozo Sugishita 2-18 Keihan Hondori, Moriguchi City, Osaka Sanyo Electric Co., Ltd.

Claims (1)

[Claims] 1. A first sound source using a pitch vector component extracted from an adaptive codebook and a code vector component extracted from a noise codebook as driving sound sources, and a code vector of the noise codebook. Driving sound source creating means for creating a second sound source that uses only components and a third sound source that uses only components of the pitch vector of the adaptive codebook, according to the first, second, and third sound sources Adaptive bit allocation means for adaptively allocating the amount of information to the pitch vector index and gain of the adaptive codebook and the code vector index and gain of the noise codebook, and obtaining the driving sound source through a weighting synthesis means. Comparing means for comparing the reproduced signal to be reproduced with the weighted input signal, and the comparing means. Comparison said first according to the result, a drive sound source determination means for determining one of the second and third sound source vector driven speech coding apparatus.