JPH0997099A - Encoding/decoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To transmit only one piece of code table code per a frame. SOLUTION: A power calculation part 1 obtains sound power at every frame, and an LPC analysis part 2 obtains an LPC coefficient, and a characteristic vector calculation part 4 obtains an LPC cepstrum coefficient. A vector quantization part 20 makes the code table code answering to a code table vector closest to a characteristic vector of a present frame different from any one of code table vectors of frame codes of past (k-1) pieces of frames among k pieces of code table vectors until k-th vector from the vector close to the characteristic vector of the present frame a frame code. A class function is obtained from the frame codes of the present frame and the past (k-1) pieces of frames. A pitch extraction part 3 obtains a pitch of a sound. A vector inverse quantization part 22 calculates the characteristic vector of the present frame, and an element piece forming part 6 forms a sound element piece from the characteristic vector. A waveform superposition part 7 superposes the sound element piece by the pitch.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency coding / decoding apparatus using vector quantization, which is capable of obtaining high-quality reproduced speech at a high compression rate.

[0002]

2. Description of the Related Art Conventionally, techniques in such a field include:
For example, some documents were described in the following documents. Reference 1: Sadahiro Furui, "Digital Speech Processing", Tokai University Press, Digital Technology Series Reference 2: IEICE Technical Report RC90-26, 1990, IAGeson
Written, `` Vector sum exited Linear prediction (VSELP) s
"peech coding for JAPAN digitalcelluar" Reference 3: ICASSP 87, 1987, HPtseng et al.,
`` Fuzzy Vector Quantization Applied to Hidden Mark
ov Modeling "Reference 4: Speech Study Group Material, 1998, Nakamura et al.," Spectral Normalization Study Using Fuzzy Vector Quantization ", sp 87-123 A vector that collectively represents a set of parameters with a single code. Quantization is described in Document 1 and is a very efficient encoding method.

In general, this method can reduce the quantization distortion as the number of representative vectors in the codebook increases, but it has a feature that the calculation amount required for the codebook search and the storage capacity of the codebook increase. The case where the vector quantization is applied to the speech coding is described in the above-mentioned Document 2. In consideration of the above characteristics, a method of interpolating with each other by using a plurality of codebooks is used. However, there are problems such as the complexity of and the increase in processing amount. The fuzzy vector quantization described in Document 3 is a method of expressing an input vector by a degree of membership in a codebook vector, and can reduce quantization distortion without increasing the number of representative vectors in the codebook.
The K-nearest neighbor rule described in Reference 4 proposed by Nakamura et al.
For each frame, fuzzy vector quantization is performed by limiting to k codebook vectors from the one closest to the input vector, and the number of power function values to be calculated is the same for all conventional codebook vectors. Since the number is reduced to k, it is extremely practical in terms of processing amount and storage capacity.

[0004]

However, when the conventional vector quantization described in Document 4 is applied to a high-efficiency speech code, k codebooks in addition to k series functions per frame are used. There is a problem that the code needs to be transmitted and the bit rate becomes high.

[0005]

In order to solve the above problems, the present invention provides a feature vector extraction unit for extracting a feature vector of input speech for each frame, and a codebook code for each frame. A vector quantizer that selects a codebook vector from a plurality of codebook vectors that it has, and uses a codebook code corresponding to the selected codebook vector as a frame code of the frame, and for each frame, the frame Based on the distance between the codebook vector represented by the code and the feature vector of that frame,
A series function calculation unit that calculates a series function value, a storage unit that stores the frame code and the series function value for each frame, and for each frame, the frame code and the class that are stored in the storage unit. In the encoding / decoding device including the vector dequantization unit that calculates the feature vector of the frame based on the function value, the vector quantization and the vector dequantization are configured as follows. .

That is, the vector quantizer is configured to code k codes from the plurality of codebook vectors to the k-th (k ≧ 2 natural number) closest to the feature vector of the current frame. From the book vector, (k
-1) Different from any of the codebook vectors represented by the frame codes of (k-1) frames up to the frame, the codebook code corresponding to the codebook vector closest to the feature vector of the current frame is set to the current frame. A frame code determining unit to be a frame code of the current frame, a codebook vector represented by each of the frame codes of the current frame and k frames up to the immediately preceding (k-1) frame, and the feature vector of the current frame. Based on each distance,
And a series function calculation unit for calculating a series function value of the current frame. The vector dequantization unit is configured to classify the codebook vector and the current frame represented by the frame code of the current frame stored in the storage unit and the k frames up to the immediately preceding (k-1) th frame. The feature vector of the current frame is calculated from the function value.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram of a speech encoding / decoding device showing an embodiment of the present invention. The voice encoding / decoding device according to the exemplary embodiment of the present invention is different from the conventional voice encoding / decoding device in that the frame code of the current frame is referred to by referring to the frame code of the immediately preceding (k−1) frame. This is because only one is determined and fuzzy vector quantization limited to the frame codes for the immediately preceding k frames including this is performed. As shown in FIG. 1, the speech encoding / decoding device according to the present embodiment includes a power calculation unit 1 and an LP.
C analysis unit 2, pitch extraction unit 3, feature vector extraction unit 4,
The vector quantization / dequantization unit 5, the segment generation unit 6, the waveform superposition unit 7, and the storage unit 8 are included. The power calculation unit 1 that calculates the power of a voice receives the sampled voice as an input. The output side of the power calculator 1 is an LP for one frame.
An LPC analysis unit 2 that performs C (Linear Predictive Coding) analysis and a storage unit 8 are connected. The output side of the LPC analysis unit 2 is connected to a pitch extraction unit 3 that calculates the pitch of the voice within one frame and a feature vector extraction unit 4 that calculates the LPC cepstrum coefficient. Pitch extractor 3
The storage unit 8 is connected to the output side of. The output side of the feature vector extraction unit 4 is connected to the vector quantization / inverse quantization unit 5. The vector quantizer / inverse quantizer 5 includes a vector quantizer 20, a codebook 21, and a vector dequantizer 22. The output side of the codebook 21 is connected to the vector quantization unit 21 and the vector dequantization unit 22. The storage unit 8 is connected to the output side of the vector quantization unit 21, and the vector dequantization unit 22 is connected to the output side of the storage unit 8.

FIG. 2 is a block diagram of the vector quantization / inverse quantization section in FIG. As shown in FIG. 2, the vector quantization unit 20 includes a code buffer 31, a k-nearest neighbor code selection unit 32, a frame code determination unit 33, and a series function calculation unit 34. On the output side of the codebook 21, a k-nearest neighbor code selection unit 32 and a class function calculation unit 34 are connected. The input side of the k-nearest neighbor code selection unit 32 is connected to the feature vector extraction unit 4 in FIG.
The frame code determination unit 33 is connected to the output side of, and the storage unit 8 and the series function calculation unit 34 are connected to the output side of the frame code determination unit 34. The storage unit 8 is connected to the output side of the series function calculation unit 34. The output side of the codebook 21 and the storage unit 8 is connected to the vector dequantization unit 22, and the output side of the vector dequantization unit 22 is
The segment creating unit 6 in FIG. 1 is connected. The code buffer 31 is a local storage area that holds the frame code for the immediately preceding k (a natural number of k ≧ 2) frames, and its value is continuously stored, but is updated once per frame as described later. . The k neighborhood code selection unit 32 uses the codebook 2
With reference to 1, the “k neighborhood code” of the “k neighborhood” is selected. The frame code determination unit 33 determines the "frame code" of the current frame.

Here, the "k neighborhood" means k codebook vectors from the closest distance to the input vector to the kth. “K neighborhood code” refers to k codebook codes corresponding to k codebook vectors,
The vector closest to the j-th is referred to as a code as X _j . The “frame code” is a storage parameter that is determined once for each frame, and the frame code of the current frame is described as f ₀ and the frame code before the i frame is described as f _i . The codebook 21 is obtained by clustering LPC spectrum coefficients extracted from a large amount of learning data in advance so that distortion is minimized by the LBG algorithm. The LPB algorithm is described in detail in Document 5 below. Reference 5: IEEE COM-28, 1980-01, Li
nde, Buzo, Gray, `` An Algorithm for Vector Quanti
zer Design ”Hereinafter, the operation of FIG. 1 will be described with reference to these drawings.

(1) Encoding processing A voice signal input from a microphone is digitized by an A / D converter after passing through an amplifier and a low-pass filter, and a frame having a fixed section length (for example, a frame length of 10 ms) is passed through a buffer. , Every 5 ms of frame shift period) is sent to the power calculation unit 1. The power calculation unit 1 calculates the power of the current frame by calculating the sum of squares of each sample of the voice signal, and outputs the voice signal to the LPC analysis unit 2 and the power to the storage unit 8, respectively. . LPC analysis unit 2
Is a speech prediction signal from a linear prediction model by the method of least squares L
The PC coefficient is determined, and the feature vector calculation unit 4 stores the LPC.
The coefficient, the LPC coefficient and the audio data for one frame are sent to the pitch extraction unit 3, respectively. The feature vector calculation unit 4 calculates the recurrence formula using the LPC coefficient,
LPC cepstrum coefficient (has a property close to human auditory characteristics, and converts an audio signal into an FFT (Fast Fourier Transfor
m), logarithmic transformation, and the same as the one obtained by inverse FFT), and using this as a feature vector, the vector quantization unit 2
Send to 0. The above LPC analysis and LPC cepstrum calculation method are described in detail in Document 1 above.

The feature vector transmitted from the feature vector extraction unit 4 in FIG. 1 (hereinafter referred to as the input vector X)
Is input to the k-nearest neighbor code selection unit 32 in FIG. At this time, k frame codes {f ₁ , f ₂ , ..., F _k } are stored in the code buffer 31. The k-nearest neighbor code selection unit 32 performs the distance between the input vector X and all the codebook vectors stored in the codebook 2, and sorts the results to obtain the k-nearest neighbor code X _j (j = 1, ...,).
k), and outputs this to the frame code determination unit 33. For example, the cepstrum distance is used as the distance measure. FIG. 3 is a flowchart showing the operation of the frame code determining unit 33 in FIG. Hereinafter, the operation of the frame code determination unit 33 will be described with reference to FIG. The frame code determination unit 33 determines in step S1 that j
Initialize to = 1. In step S2, the first neighborhood code X ₁ is compared with the (k−1) frame codes {f ₁ , ..., F _k−1 } on the code buffer 31. No comparison is made with the frame code f _k before k frames.

In step S3, the first neighborhood code X
_{If 1} is different from any of the (k-1) codes in the code buffer 31, the frame code f ₀ is determined to be the first neighbor code X ₁ , and the process proceeds to step S4, where (k -1) If it is equal to any one of the frame codes, go to step S3. In step S3, j is incremented by 1, and the process returns to step S2. In step S2, this time the second neighborhood code X
₂ and (k-1) codes {f on the code buffer 31
_{, 1} , ..., F _k−1 } are compared. In step S3,
The second neighborhood code X ₂ is (k−
1) If it is different from any of the code, the frame code f ₀ is determined as the second neighborhood code X ₂ , and step S4
If it is equal to any one of the (k-1) codes on the code buffer 31, then j is incremented by 1 again in step S3 and the process returns to step S2 again, this time the third neighborhood. Code X ₂ and code buffer 3
(K-1) frame codes {f ₁ , ..., F on _1
k-1 } are compared. If such a procedure is repeated at most for the k-nearest neighbor code X _k , the frame code X _j can be determined and the process can proceed to step S4. The frame code X _j determined in this way differs from any of the (k−1) frame codes in the code buffer 31, and is the most significant of the input vector X among the codebook vectors stored in the codebook 21. It is a code corresponding to a codebook vector having a short distance.

In step S4, the frame code X
Let _j be the frame code f ₀ . In step S5, the frame code f ₀ is output to the storage unit 8, and the process proceeds to step S6. In step S6, {f ₁ , f ₂ , ..., F _k } stored in the storage unit 8 are changed to {f ₀ , f ₁ , ..., F.
_k-1 }. Here, the obtained frame codes {f ₀ , f ₁ , ..., F _k-1 } are all different from each other, and the first neighborhood X ₁ has the frame code {f ₀ , f ₁ , ..., F.
Always included in _k-1 }. Moreover, since the voice has a characteristic that it continuously changes, when the voice changes smoothly, the frame code {f ₀ , f ₁ , ..., F _k-1 }
Is a set of k neighborhood codes {X ₁ , X ₂ , ...,
Many if it matches _{X k}, {f 0,} f 1, ..., f
It can be said that the k codebook codes of ( _k-1 ) are close to k. The series function calculation unit 34 recalculates the distance between the codebook vector indicated by each of the k codes {f ₀ , ..., F _k−1 } on the updated code buffer 31 and the input vector X, Quantize or vector quantize to an appropriate bit using this, and output it to the storage unit 8 as a series function value, or calculate only the class function value λ 0 corresponding to the nearest code of the input vector X. , And may be quantized and output to the storage unit 8 together with the pointer code that points to the nearest neighbor.
However, in the latter case, processing such as uniformly giving each class function value corresponding to another neighborhood code as (1-λ0) / (k-1) is added to the vector dequantization unit 22. It is necessary to keep it. The method of obtaining the class function value is described in detail in Document 5 below.

Reference 5: Japanese Patent Application Laid-Open No. 1-237600 On the other hand, the pitch extraction unit 3 in FIG. 1 uses a modified correlation method (L) which uses an LPC residual waveform calculated from an LPC coefficient and an audio signal.
When the PC residual waveform is shifted by the bitch, the phases almost match,
The power before and after the shift is maximized) is used to determine the bitch length, and the bitch length is sent to the storage unit 8. The modified correlation method using the LPC residual waveform is described in Document 6 below. Reference 6: Reports of the Int.Con.Acoust.Tokyo, C-5-5
, 1968, edited by Y.kohashi, F. Itakura, S, Saito
Author, `` Analysis Synthesis Telephonybased on the Max
imum Likelihood Method ”(2) Decoding Processing Next, the vector dequantization unit 22 will be described. The frame code received from the storage unit 8 is used to update a code buffer (not shown) in the vector dequantization unit 22.
The procedure is the same as that of the code buffer 31 in the vector quantizer 20. That is, from k codes {f ₁ , f ₂ , ..., F _k } on the code buffer to {f ₀ , f ₁ , ..., f
_k-1 } is updated. Next, k received from the storage unit 8
The codebook vectors corresponding to the _k codes {f ₀ , f ₁ , ... F _k-1 } on the code buffer are obtained from the codebook 21 using the series function values as coefficients, and those codebook vectors are obtained. By linearly combining, a feature vector that is an LPC cepstrum coefficient is obtained and sent to the segment creating unit 6 in FIG. The segment generation unit 6 sequentially performs FFT, exponential conversion, and inverse FFT on the LPC cepstrum coefficient to generate a symmetric speech segment (LP
The C cepstrum coefficient is created by using only the real number part of the FFT, and the imaginary number part is discarded, so that the voice segment is symmetric) is created and sent to the waveform superimposing part 7.

The waveform superposing unit 7 determines the power of the speech unit from the pitch and the power received from the storage unit 8,
After the unit power is adjusted, the units are superimposed at a pitch interval to synthesize a voice and output from the speaker via the output buffer and the D / A converter. The waveform superimposing unit 7
The power adjustment of the element and the method of superimposing the element in (3) are described in detail in Document 6 below. Reference 6: Acoustical Society of Japan, 44, 11 , 1988, Nakajima et al., "Power spectrum envelope (PSE) speech analysis."
Synthesizing System ”As described above, according to the embodiment, there is an advantage that the fuzzy vector quantization for accumulating only one code per frame can be realized and the storage capacity of the voice accumulating device can be significantly reduced.

The present invention is not limited to the above embodiment, and various modifications can be made. For example, there are the following modifications. (1) In the present embodiment, the LPC cepstrum is used as the characteristic parameter, but another parameter such as the improved cepstrum coefficient may be used. (2) Further, the present invention can be applied to all speech coding methods in which vector quantization is performed for each frame, and its usage form is not limited to the waveform superposition type speech storage device. (3) Furthermore, the present invention can be used for media other than voice, for example, data compression or encoding of images and the like.

[0017]

As described in detail above, according to the present invention, vector quantization for accumulating only one code per frame can be realized, and the storage capacity of the storage unit can be greatly reduced.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a speech encoding / decoding device according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a vector quantization / inverse quantization unit in FIG.

3 is a flowchart showing an operation of a frame code determination unit in FIG.

[Explanation of symbols]

 1 Power Calculation Section 2 LPC Analysis Section 3 Pitch Extraction Section 4 Feature Vector Extraction Section 5 Vector Quantization / Inverse Quantization Section 6 Element Creation Section 7 Waveform Superposition Section 8 Storage Section 20 Vector Quantization Section 21 Codebook 22 Vector Inverse Quantization Conversion unit 31 code buffer 32 k neighborhood code selection unit 33 frame code determination unit 34 class function calculation unit

Claims (1)

[Claims] 1. A feature vector extraction unit for extracting a feature vector of input speech for each frame, and for each frame, a codebook vector is selected from a plurality of codebook vectors each having a codebook code. Then, for each frame, a vector quantizer that uses the codebook code corresponding to the selected codebook vector as the frame code of the frame, and for each frame, the distance between the codebook vector represented by the frame code and the feature vector of the frame. Based on a series function calculation unit for calculating a series function value, a storage unit for storing the frame code and the series function value for each frame, and for each frame, the frame code stored in the storage unit. And a vector dequantization unit that calculates a feature vector of the frame based on the class function value. In the encoding / decoding device, the vector quantizing unit includes k number of codebook vectors from the closest distance to the feature vector of the current frame to the k-th (k ≧ 2 natural number). , Which is different from any of the codebook vectors represented by the frame codes of (k-1) frames up to the immediately preceding (k-1) frame, the distance from the feature vector of the current frame is the largest. A frame code determination unit that uses a codebook code corresponding to a close codebook vector as a frame code of the current frame, and the frame code of each of k frames up to the (k-1) th frame immediately before the current frame. A class function calculator that calculates a class function value of the current frame based on each distance between the codebook vector and the feature vector of the current frame; Le inverse quantization unit, the current frame stored in the storage unit and the immediately preceding (k-
1) A coding / decoding device characterized in that a feature vector of a current frame is calculated from a codebook vector represented by each of frame codes of k frames up to a frame and a series function value of the current frame. .