JPH0527799A - Method and device for vector quantization - Google Patents

Abstract

PURPOSE:To suppress the generation of an abnormal sound by suppressing the discontinuity of a spectrum between frames. CONSTITUTION:The logarithmic power spectrum regeneration part of a vector encoding system reception-side system consists of a spectrum conversion part 501 which composes a logarithmic spectrum of a code spectrum 114, a power comparison part 502 which switches subsequent processes by comparison with the power of a last frame, a frequency distribution part 503 which discriminates between the common component 504 and other component 505 of the logarithmic spectrum, a weighted mean part 506 which averages the common part 504 with the frequency component 508 of the last frame, a frequency mixing part 510 which mixes the weighted mean component 509 with the other discriminated frequency component 505 again, and a level regeneration part 512 which adds level information 105'.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency speech coding apparatus, and more particularly to a vector quantization method and apparatus suitable for obtaining high-quality reproduced speech at a high information compression rate.

[0002]

2. Description of the Related Art Conventionally, various systems have been proposed as a voice efficiency coding system. For example, Kazuo Nakata "Digital Information Compression" (Kosaido Kogyo Publishing, Electronic Science Series 10
In 0), various methods are explained in an easy-to-understand manner.
A number of schemes relating to waveform coding schemes and information source coding schemes (parameter coding schemes) are shown.

[0003]

Among these various methods, the waveform coding method has good sound quality, but it is difficult to increase the information compression efficiency, and the parameter coding method has high information compression efficiency. On the contrary, even if the amount of information is increased, the sound quality has an upper limit, and there is a drawback that sufficient quality cannot be obtained. In particular, information compression in the middle of the band that both parties are good at (10
Near kbps) is a valley band. On the other hand, as a hybrid method combining the advantages of both methods, a multi-pulse method (for example, BSAtal et al. “A new mo
del of LPCexcitation for producing natural-soundin
g speech at low bit rates ”Proc.ICASSP 82, S-
5, 10, (1982), etc., and the TOR method (A.Ichi
kawaet al., “A speech coding method using thinned-
out residual "Proc.ICASSP85, 25.7 (198
Although 5)) and the like have been proposed in recent years and various studies have been made, they are inadequate in terms of sound quality and cost of processing. In general, various high-efficiency coding schemes are realized by paying attention to the fact that the existence of voice information is biased and thickening the allocation of codes to the parts where the information exists. We will further promote this point,
Paying attention to the bias of information as a combination of a plurality of parameters, the method of thickening the code allocation to the part where the speech information exists for the combination set of parameters (called a vector) (called vector quantization) ( For example, S. Roucos
et al., “Segment quantization for very-low-rate s
peechcoding "Proc.ICASSP 82, p. 1563 (19)
82)) is drawing attention.

In vector quantization, a codebook that stores a code vector (also called parameter, representative vector, or codeword) in association with a code (also called index or index) is used. Vector quantization is a quantization process in the sense that an input vector is mapped to a finite number of code vectors, but since the amount of information of the index can be made smaller than the original vector, it is used for information compression.
CE applying vector quantization to prediction residual of speech
LP method (eg BSAtal et al., “Stochastic cod
ing of speech signals at very low bit rates ”Proc.
ICC84, pp. 1610-1613 (1984)) and its improved VSELP method (for example, IAGerson).
et al., ”Vector sum excited linear prediction (VSE
LP) ”Proc.IEEE workshop on speech coding for telec
ommunications, pp. 66-68 (1989)) and the like. These methods employ a method that minimizes the error between the synthetic speech and the original speech by using a feedback loop. Therefore, the quality of synthetic speech is good,
There is a problem that the processing amount is large and that it is difficult to apply to 4.8 kbps or less. In order to realize high quality speech coding, it is necessary to create a good quality codebook in advance. It is fundamental to reduce the quantization distortion to improve the quantization characteristics, but the quantization distortion becomes smaller as the codebook becomes larger. Also, as one of the problems caused by performing vector quantization using a codebook of a finite number of sizes, if there is a large component of discontinuity in the characteristics between adjacent speech frames, there will be abnormal noise in the reproduced speech. May occur. In order to reduce the discontinuity of characteristics between frames, it is necessary to increase the code vectors included in the codebook and reduce the distance between each code vector, but this requires setting a codebook of a considerably large size. Becomes However, in order to create a large-scale codebook, an extremely large amount of voice data must be used, and a huge amount of processing is required. Further, the size of the codebook is limited due to restrictions such as memory capacity and search processing amount.

It is an object of the present invention to suppress a change in spectrum between two consecutive frames within a certain degree in a vector quantization method and its apparatus, and thereby cause a difference caused by a drastic change in spectrum characteristics between frames. It is to suppress the generation of sound and improve the quality of reproduced voice.

[0006]

In order to achieve the above object, in a vector quantization method and its apparatus, a band in which the degree of change in spectrum is remarkable between two consecutive frames in which the change in power is not steep is input. Focusing on the fact that the characteristics of speech are not reflected so much, the spectrum in that band is averaged with the spectrum of the previous frame with weighted averaging to make the spectrum common to some extent, and It is configured to suppress the change within a certain amount.

[0007]

Various modifications of the present invention are conceivable, and the operation of typical means will be described below.

When the voice to be transmitted is input, the feature vector is extracted in the analysis unit and sequentially compared with the code vector in the codebook, and the closest vector is selected. At this time, the distance between the code vector selected in the previous frame and the code vector selected in the current frame increases as the size of the codebook decreases unless the same code vector is selected in two frames. .

Therefore, first, the spectrum information in the previous frame is stored for the portion where the power does not change suddenly between frames. Next, after restoring the spectrum for the code vector selected in the current frame, weighted averaging is performed with respect to the above-mentioned spectrum stored for the frequency region where the spectrum discontinuity is large and abnormal noise is remarkable. I do.

In the part where the power for performing the averaging process does not change suddenly, the characteristic originally possessed by the voice is hardly separated so much between consecutive frames, and it is selected by this method. Since the spectrum is averaged only in the frequency band where abnormal noise is generated with respect to the code vector and no processing is performed in other bands, even if the spectrum is modified by this method. There is almost no loss of clarity of the reproduced voice. Further, since the processing of this method is entirely performed only on the receiving side, the information transmitted from the transmitting side remains the same as before and the transmission amount does not increase.

[0011]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram for explaining an embodiment of the present invention. Only one direction in which the transmitting side and the receiving side are paired is shown, and the communication path in the opposite direction is omitted because the figure becomes complicated.

In FIG. 1, an input voice 101 is input to a two-sided buffer memory 103 via an analog / digital (A / D) converter 102. This memory is provided to adjust the time of the following processing and prevent interruption of the input voice. The voice from the buffer memory 103 is input to the analysis unit 104, and the level information 105, the spectrum information 106, and the pitch information 107 are obtained. The spectral information 106 is added to the vector quantizer 108,
Obtain the vector code 109. Vector code 109,
The pitch information 107 and the level information 105 are transmitted by the transmitter 110,
It is sent to the receiving unit 112 via the transmission path 111. On the receiving side, the vector code 109 ', the pitch information 107', and the level information 105 'received by the receiving unit 112 are added to the vector dequantization unit 113, the spectrum information 114 is restored, and the pitch information 107' and the level information 105 '. It is added to the combining unit 115 together with. The synthesizing unit 115 decodes it into a voice waveform, passes through the output two-sided buffer memory 116,
A digital-to-analog (D / A) converter 117 converts the analog signal and reproduces it as an output sound 118.

Each part will be described in detail below.

FIG. 2 is a diagram for explaining the analysis unit 104. In the present embodiment, the analysis unit uses the power spectrum envelope (P
SE) Analytical method. The PSE analysis method is described in detail in Nakashima et al.'S paper "Power Spectrum Envelope (PSE) Speech Analysis / Synthesis System", Journal of Acoustical Society of Japan, Vol. 44, No. 11 (Sho 63-11). The outline is given here.

In FIG. 2, a pitch extraction unit 201 extracts pitch information (pitch frequency or pitch period) of input speech. A known method such as the correlation method or the AMDF method may be used as the pitch extraction method. The waveform cutout unit 202 cuts out a waveform section for analyzing spectrum information from the input voice, and cuts out a section of about 20 to 60 ms. It is often a fixed-length section, but it may be a variable length that is about three times as long as it depends on the pitch period. The cut out waveform is sent to the Fourier transform unit 203 and converted into a Fourier series. At this time, after multiplying the cut-out waveform by a window function that is normally used, such as a Hamming window, embedding zero data before and after to obtain 2048-point data, and using fast Fourier transform (FFT), high speed and frequency resolution High data is obtained. The Fourier coefficient displayed as an absolute value is the frequency component of the cut-out waveform, that is, the spectrum. When the cutout waveform has a periodic structure, the spectrum has a line spectral structure due to pitch harmonics.

The pitch resampling unit 204 extracts only the harmonic component (line spectrum component) of the pitch frequency from the spectrum information obtained by the FFT. The data thus extracted will be considered below in association with the period π when the cosine series is expanded, which will be described later.

The power spectrum conversion section 205 squares each component of the spectrum and converts it into a power spectrum. Further, the logarithmic unit 206 logarithmizes each component to obtain a logarithmic power spectrum.

The level normalization unit 207 absorbs level fluctuations based on the volume of the input voice, but may be collectively extracted in the next cosine transform unit 208.

The cosine transform section 208 uses data obtained by re-sampling the logarithmic power spectrum to approximately express it as a finite term cosine series. The number of terms m is usually set to about 25. The power spectrum envelope is expressed as follows.

[0021]

[Equation 1] Y = A0 + A0 · A1 · cosλ + A0 · A2 · cos2λ + ... + A0 · Am · cos mλ The coefficient A is such that the square error between the resampled power spectrum data and Y according to Equation 1 is minimized. Required to. Since the 0th term A0 of the coefficient represents the input level, the level information 105 and A1, ..., Am are output as the spectrum information 106.

Next, the vector quantizer will be described with reference to FIG. In FIG. 3, a codebook 301 stores values of code vector elements and their codes. In the matching unit 302, the spectrum information (input vector) 1
When 06 is input, each code vector is read from the code book 301 and the distance from the input vector 106 is calculated. Here, the distance measure is the Euclidean distance in which each element of the vector is weighted, but it goes without saying that another suitable measure may be used. In the case of exhaustive search, distance values are calculated for all code vectors in the codebook, and the code 109 of the code vector (nearest neighbor vector) that minimizes the distance value is output. Note that it is possible to limit the range of code vectors to be matched by using the pitch information 107 and the like. As a vector quantization method, complementary vector quantization (Japanese Patent Application No.
It is also possible to adopt a method such as 3-75700).

The codebook 301 will be described below. The training data is processed in the same way as in the analysis unit of FIG. 1 to obtain a training vector, and a plurality of training vectors distributed in the signal feature space are classified into a plurality of clusters smaller than the number of training vectors. Create a representative vector from the cluster and register it in the codebook. As a method of creating the representative vector, first, the center of gravity is obtained in each cluster, and the training vector having the smallest distance from the center of gravity of the training vector distributed in each cluster is used as the representative vector of the cluster. Alternatively, there is a method in which the center of gravity obtained by using a plurality of training vectors whose distance from the center of gravity of the cluster is included in a predetermined range is used as the representative vector. In this embodiment, about 5000 training vectors are used and 10
A codebook of 24 sizes was created.

Next, the decoding side (reception side) will be described.

In FIG. 1, when the vector dequantization unit 113 receives the vector code 109 ', the corresponding code vector is read from the codebook. In addition,
It goes without saying that the codebook on the receiving side has the same contents as the codebook 301 on the transmitting side. The reproduction code vector y ′ = {A1 ′, A2 ′, ..., Am ′} is sent to the synthesizing unit 115 as spectrum information 114.

Next, the synthesizing unit 115 will be described with reference to FIG. In the figure, a logarithmic power spectrum reproducing unit 4
01 will be described with reference to FIG. The spectrum conversion unit 501 obtains a logarithmic power spectrum Y'by using the elements A1 ', A2', ..., Am 'of the transmitted reproduction vector (spectral information 114) according to the following equation.

[0027]

## EQU2 ## Y '= 1 + A1'cosλ + A2'cos2λ + ... + Am'cos mλ Next, the power comparing section 502 compares the power with the previous frame, and the level reproducing section is applied to a frame in which power is suddenly changed such as a burst sound. The output is output to 512, and the output is switched to the frequency discriminating unit 503 in the frame where the power does not suddenly change. In the frequency discriminating unit 503, the input logarithmic power spectrum 502 is frequency component Fn for commonizing the spectrum.
504 and the other components 505 are separated. In this embodiment, the frequency band of 2 kHz or higher is shared. The weighted averaging unit 506 takes out the common component Fn-1 508 of the spectrum of the previous frame stored in the data storage unit 507 in advance, and calculates it with the common component Fn ′ 504 of the input spectrum according to Formula 3. Performs weighted averaging.

[0028]

## EQU00003 ## Fn = (W1.Fn '+ W2.Fn-1) / (W1 + W2) The respective weights W1 and W2 can be changed depending on the frequency. Output 5 of weighted averaging unit 506
09 updates the data in the data storage unit 507 and
Remaining frequency component 5 output from the frequency discriminating unit 503
05 is mixed again with the frequency mixing unit 510. The spectrum 511 output from the frequency mixing unit 510 is
Level information A0 105 'in the level reproduction unit 512
Is used, level information based on the volume of the input voice is added.

Returning again to FIG. 4, the reproduced logarithmic power spectrum 402 undergoes exponential conversion in the exponential conversion unit 403,
The zero-phased spectrum 404 is obtained, and the inverse Fourier transform unit 40 is obtained.
Sent to 5. The inverse Fourier transform unit 405 obtains the speech unit 406 by inverse fast Fourier transform (IFFT). The speech unit 406 uses the waveform synthesizer 407 to generate pitch information 1
According to 07 ', they are sequentially added while being shifted by a pitch interval, and output as reproduced voice 408.

The effect of the invention in the embodiment is that the generation of abnormal noise due to the temporal discontinuity of the spectrum of the synthesized voice is suppressed because the spectrum of the frequency of 2 kHz or more is weighted and averaged with the spectrum of the previous frame. On the other hand, 2
Since no processing was applied to frequencies below kHz, the sound quality was improved without impairing the clarity.
Regarding the amount of information transmitted, the amount of information transmitted is the same as before because the transmitting side uses the conventional vector quantization system as it is.

[0031]

According to the present invention, when the same codebook is used, the generation of abnormal noise due to the temporal discontinuity of the spectrum of the synthesized speech without deteriorating the clarity as compared with the conventional vector quantization. It is possible to improve the sound quality by suppressing the noise.

In the description of the present invention, all the objects are voices, but it goes without saying that they can be applied to those having a similar structure.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a system configuration of an embodiment of the present invention.

FIG. 2 is a diagram illustrating an analysis unit.

FIG. 3 is a diagram illustrating a vector quantization unit.

FIG. 4 is a diagram illustrating a combining unit.

FIG. 5 is a diagram illustrating a logarithmic power spectrum reproducing unit.

[Explanation of symbols]

101 ... Input voice, 103, 116 ... Buffer memory,
104 ... Analysis unit, 106, 114 ... Spectral information, 10
7, 107 '... Pitch information, 108 ... Vector quantizer, 109, 109' ... Vector code, 113 ... Vector dequantizer, 115 ... Synthesizer, 119 ... Output speech.

Claims (14)

[Claims] 1. A decoding side has a codebook in which different codes and a code vector representing a feature space of a signal corresponding to the code are stored, and by the code transmitted from the encoding side for each frame, When a code vector corresponding to a code is selected and the selected code vector is restored to a characteristic amount such as a spectrum of a signal, the previous frame is applied to a portion of the characteristic amount that does not reflect the signal characteristic. The vector quantization method, characterized in that the feature quantity is modified by performing commonization with the portion used in (1). 2. The decoding side uses the power information transmitted from the encoding side, and modifies the feature amount only for the frames in which the change in the power information between the frames is small. The vector quantization method according to claim 1, characterized in that 3. When transforming the feature quantity, weighted averaging is performed on the spectrum component restored from the codebook with the spectrum used in the previous frame. 1 vector quantization method. 4. The vector according to claim 3, wherein, when performing weighted averaging between the spectra, the spectrum component for which the weighted averaging is performed is limited to a predetermined frequency region and the averaging is performed. Quantization method. 5. The vector quantization method according to claim 3, wherein in performing the weighted averaging between the spectra, the weight used in the weighted averaging is changed for each frequency component. 6. When creating the codebook, the feature space of the signal is divided into a plurality of clusters smaller than the number of training vectors by clustering a plurality of training vectors distributed in the feature space of the signal. The vector quantization method according to claim 1, wherein a training vector having a minimum distance from the center of gravity of each cluster is adopted as the representative vector among the training vectors. 7. When creating the codebook, the feature space of the signal is divided into a plurality of clusters smaller than the number of training vectors by clustering a plurality of training vectors distributed in the feature space of the signal, The vector quantization method according to claim 1, wherein a representative vector is created using a plurality of training vectors whose distances from the center of gravity of each cluster are within a predetermined range among the training vectors. 8. A codebook creating method for creating a codebook that stores different codes and a code vector representing a feature space of a signal corresponding to the code on the decoding side, wherein the feature space of the signal is the feature space of the signal. Is divided into a plurality of clusters smaller than the number of the training vectors by clustering the plurality of training vectors distributed in the, and the training vector having the smallest distance from the center of gravity of each cluster among the training vectors is the representative vector. A codebook creation method characterized by being adopted as. 9. A codebook creating method for creating a codebook for storing different codes and a code vector representing a feature space of a signal corresponding to the code on the decoding side, wherein the feature space of the signal is the feature space of the signal. Are divided into a plurality of clusters smaller than the number of the training vectors by clustering a plurality of training vectors distributed in the above, and the distance from the center of gravity of each cluster in the training vectors is included within a predetermined range. A codebook creating method characterized in that a representative vector is created using a plurality of training vectors. 10. The decoding side has a codebook for storing different codes and a code vector representing a feature space of a signal corresponding to the code, and the codebook is transmitted by the coding side for each frame. Means for selecting a code vector corresponding to a code, and in restoring the selected code vector to a characteristic amount such as a spectrum of a signal, for a portion of the characteristic amount that does not reflect the characteristic of the signal A vector quantizer having a means for transforming the feature quantity by performing commonization with the portion used in the previous frame. 11. The decoding side uses the power information transmitted from the encoding side, and modifies the feature amount only for the frames in which the change in the power information between the frames is small. 11. The vector quantizer of claim 10, wherein the vector quantizer is a quantizer. 12. The feature amount transforming means includes means for performing weighted averaging on the spectrum component restored from the codebook with the spectrum used in the previous frame. The vector quantization device according to claim 10. 13. The vector quantum according to claim 12, wherein the weighted averaging means between the spectra performs averaging by limiting the spectrum component for which the weighted averaging is performed to a predetermined frequency domain. Device. 14. The vector quantization device according to claim 12, wherein the weighted averaging means between the spectra changes a weight for performing the weighted averaging for each frequency component.