JPH05232994A - Statistical code book - Google Patents

Abstract

PURPOSE:To attain a statistical code book capable of obtaining excellent reproduced sound quality even in a low bit rate. CONSTITUTION:The statistical code book 70 is provided with two 1st and 2nd statistical code book parts 71, 72 capable of simultaneously outputting plural statistical code vectors CBa, CBb. Both outputted code vectors weighted and added by respective weighted addition parts (73 to 75) and outputted as an output vector of the statistical code book 70. A weighting coefficient determining part 76 determines weighting coefficients alpha, 1-alpha based upon voice vectors already synthesized based upon the outputted stastistical code vector or the information of the input voice vector. Consequently reproduced sound quality similar to a case using many statistical code vectors can be obtained by the use of the smaller number of code vectors and information to be transmitted/ received by a coder and a decoder can be reduced.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical codebook provided in a code-excited linear predictive encoder or decoder.

[0002]

2. Description of the Related Art Conventionally, there is a backward type code excitation linear predictive encoder and decoder as described in the following document.

Reference: “Juin-Hwey Chen,“ High-Quality
16KB / S SPEECH CODING WITH AONE-WAY DELAY LESS THA
N 2MS, ”Proc.IEEE Int.Conf.Acoust., Speech, SingleP
rocessing.pp453-456 (1990). ”First, the backward code excitation linear predictive encoder will be described with reference to the block diagram of FIG.

In FIG. 2, the information transmitted to the receiving side is only the index of any statistical code vector stored in the statistical codebook 11. Such an optimal index is determined as follows.

At the time of searching for the optimum index, each statistical code vector stored in the statistical code book 11 is output in time sequence as a candidate.

The statistical code vector e j as a candidate output from the statistical code book 11 in this way is given to the gain circuit 12, multiplied by a predetermined (g) times by the gain circuit 12, and given to the adder 13. The gain circuit 12 is of a variable gain coefficient type, and the gain coefficient g is changed by the gain control circuit 14. The gain control circuit 14 uses the past vector series e from the gain circuit 12.
A gain coefficient g is predicted by performing a linear prediction analysis (LPC analysis) from x and is given to the gain circuit 12.

An output signal from the synthesis filter (linear prediction filter) 15 is also given to the adder 13, and the candidate statistical code vector ex from the gain circuit 12 and the output vector from the synthesis filter 15 are added, The synthesized voice vector Sw for local reproduction when the statistical code vector of the candidate is used is obtained and given to the subtractor 16.

Vocal tract prediction coefficient a used by the synthesis filter 15
i is given from the vocal tract analysis circuit (linear prediction analysis circuit) 17. The synthesis filter 15 performs a predictive synthesis process by applying the vocal tract prediction coefficient ai given from the vocal tract analysis circuit 17 to a series of locally reproduced synthetic speech vectors with respect to the past optimum statistical code vector, and outputs the output vector. To the adder 13. The vocal tract analysis circuit 17 obtains a vocal tract prediction coefficient a i from a series of synthesized speech vectors for the optimum statistical code vector in the past and gives it to the synthesis filter 15.

An input speech vector S is also given to the subtractor 16, and the subtractor 16 calculates from the input speech vector S
The synthesized voice vector Sw for local reproduction when the candidate statistical code vector ej is used is subtracted, and the obtained difference vector er is weighted according to the auditory characteristic through the auditory weighting filter 18, and then the index search is performed. It is given to the circuit 19. In this way, at the time of searching for the optimum statistical code vector, the index search circuit 19 is provided with the difference vector ew which is auditory-weighted for all statistical code vectors.

The index search circuit 19 calculates the sum of squares of the components of each difference vector ew, and detects the statistical code vector corresponding to the difference vector ew having the minimum square sum as the optimum statistical code vector. The index I is assigned to the quantizer 20 and the statistical codebook 11.
Give to.

The quantizer 20 codes the index I of the optimum statistical code vector into a code (C) and transmits it to the receiving side. The statistical codebook 11 outputs the detected optimum statistical code vector to the gain circuit 12 side again to update the gain coefficient and the linear prediction coefficient.

Next, referring to the block diagram of FIG. 3, FIG.
A conventional backward type code excitation linear predictive decoder corresponding to the above encoder will be described.

In FIG. 3, the transmitted code C is input to the inverse quantizer 21, converted into the index I, and given to the statistical codebook 22. The statistical codebook 22 has the same configuration as the statistical codebook 11 on the encoder side, and outputs the statistical code vector e related to the given index I to the gain circuit 23. This statistical code vector is given a predetermined value (g) by the gain circuit 23.
It is multiplied and given to the adder 24. The gain circuit 23 is also a variable gain coefficient type, and the gain coefficient g is changed by the gain control circuit 25. Gain control circuit 25
Is a linear prediction analysis (LPC analysis) from the past vector series from the gain circuit 23 to predict the gain coefficient g and gives it to the gain circuit 23.

An output vector from a synthesis filter (linear prediction filter) 26 is also given to the adder 24. The statistical code vector ex from the gain circuit 23 and the output vector from the synthesis filter 26 are added, and the statistics are obtained. A synthesized voice vector Sw for reproduction corresponding to the code vector is obtained and output from the output terminal.

Vocal tract prediction coefficient a used by the synthesis filter 26.
i is given from the vocal tract analysis circuit (linear prediction analysis circuit) 27. The synthesis filter 26 performs a prediction synthesis process by applying the vocal tract prediction coefficient ai given from the vocal tract analysis circuit 27 to the sequence of the reproduced synthetic speech vector Sw for the past statistical code vector, and outputs the output vector. It is given to the adder 24. The vocal tract analysis circuit 27 obtains a vocal tract prediction coefficient ai from the sequence of the synthetic speech vector Sw for the past statistical code vector and gives it to the synthesizing filter 26.

[0016]

However, in the above-mentioned conventional method, since the excitation signal is represented by one statistical codebook 11 or 22 for all input voices, the prepared statistical codebook is used. It has a drawback that the synthesized voice quality is deteriorated with respect to the incompatible voice. For example, when a statistical codebook that considers male voices is prepared, the playback sound quality deteriorates with respect to female voices. There has been a problem that the reproduced sound quality for voice has deteriorated.

Recently, encoding at a low bit rate has been required, but when the low bit rate is expected, the number of statistical code vectors stored in the statistical code book becomes small. Therefore, it is practically difficult to store a plurality of types of statistical code vectors (male and female voices, Japanese voices and foreign language voices) in one statistical codebook, and the above-mentioned drawbacks are great. ..

The present invention has been made in consideration of the above points, and it is an object of the present invention to provide a statistical codebook which can obtain excellent synthesized speech quality even at a low bit rate.

[0019]

In order to solve such a problem, in the present invention, the statistical codebook provided in the code-excited linear predictive encoder or decoder is configured by the following units.

That is, a plurality of statistical codebook sections storing different statistical codevectors and a plurality of statistical codevectors output from each statistical codebook section are weighted and added to form a statistical codevector to be output. The addition unit and the coefficient determination unit that determines the weighting coefficient used by the weighted addition unit based on the information of the voice vector or the input voice vector already synthesized based on the output statistical code vector.

[0021]

In the present invention, a plurality of statistical codebook sections for simultaneously outputting statistical code vectors are provided. The plurality of statistical code vectors output from these statistical codebook units are weighted and added by the weighted addition unit and become the output vector of the statistical codebook. here,
The weighting coefficient is determined by the coefficient determining unit based on the information of the voice vector or the input voice vector already synthesized based on the output statistical code vector. As a result, it becomes possible to obtain the quality of the reproduced speech similar to the case where a larger number of statistical code vectors is used with a smaller number of statistical code vectors, and even in that case, the encoder and the decoder exchange data. Less information is required.

[0022]

【Example】

(A) First embodiment (A-1) Code-excited linear predictive encoder to which the statistical codebook of the first embodiment is applied Before describing the statistical codebook of the first embodiment, first,
A code excitation linear prediction encoder to which this is applied will be described with reference to FIG.

In FIG. 4, the original audio signals are grouped in units of frames and the input audio vector S is input terminal 3
Input to 1. A plurality of adaptive code vectors Vaj (j = 1 to n) and a plurality of statistical code vectors Vsk stored in the adaptive codebook 32 and the statistical codebook 33.
From (k = 1 to m), the optimum adaptive code vector Va and the statistical code vector Vs for the input speech vector S at the current time are searched as follows, and the searched optimum adaptive code vector Va and the statistical value are searched. The indexes Ia and Is of the code vector Vs are given to the quantizer 34.
The quantizer 34 collects these indexes Ia and Is into a total code C and outputs it from the output terminal 35.
The statistical code book 33 has the specific structure shown in FIG. 1, and details thereof will be described later.

The search is performed in the following order. First, the optimum adaptive code vector Va is searched while the statistical code vector is not output, and then the output of the searched optimum adaptive code vector Va is fixed to search for the optimum statistical code vector Vs. ..

When searching for the optimum adaptive code vector Va, the adaptive code book 32 supplies a plurality of adaptive code vectors Vaj as candidates to the thinning circuit 36 in time sequence or simultaneously. The thinning circuit 36 thins out the components of each adaptive code vector Vaj to convert into a vector Vdj in which the number of components is 1 / x times (x is an arbitrary number) and supplies it to the adder 37.

The adder 37 is also provided with the statistical code vector from the statistical code book 33, but is not provided at the time of searching for the optimum adaptive code vector Va. At this time, the vector Vdj becomes the excitation code vector Vj which is a combination of the adaptive code vector and the statistical code vector. The excitation code vector Vj is given to the multiplier 39.

The multiplier 39 outputs the excitation code vector Vj
Is multiplied by the excitation gain g given from the gain control circuit 40, and the vector Vgj after multiplication is given to the adder 41.
The gain control circuit 40 applies, for example, a linear prediction analysis to the past vector Vg sequence to predict the excitation gain g at the current time and supplies it to the multiplier 39.

The output vector Sp from the synthesis filter 42 is also given to the adder 41, and the vector Vgj and the vector Sp are added to add the synthesized voice vector S for local reproduction.
sj is obtained and given to the subtractor 43. A vocal tract analysis circuit (for example, a linear prediction analysis circuit or an IMEM analysis circuit) 44 obtains a vocal tract prediction coefficient ai at the current time by using the past synthesized speech vector Ss sequence that has already been processed and obtains the synthesis filter 42.
Give to. The synthesizing filter 42 synthesizes the past synthesized speech vector Ss sequence using the vocal tract prediction coefficient ai at the present time to obtain the above-mentioned vector Sp.

The subtractor 43 obtains a difference vector ej between the input speech vector S and the synthesized speech vector Ssj corresponding to the candidate adaptive code vector Vaj and gives it to the perceptual weighting filter 45. The perceptual weighting filter 45 performs a weighting process on the difference vector ej according to the perceptual characteristics, and supplies the output vector ewj to the index search circuit 46.

The index search circuit 46 uses the vector e
The root mean square of each component of wj is calculated, the adaptive code vector Va having the smallest value is detected, and its index Ia is detected.
To the quantizer 34 and the adaptive codebook 32.

Next, the process proceeds to search for the optimum statistical code vector Vs. The adaptive codebook 33 has index I
The optimum adaptive code vector Va relating to a is thinned out by the thinning circuit 3
6, and the thinning circuit 36 supplies a vector Vd obtained by thinning a predetermined component of the vector Va to the adder 37. To the adder 37, the statistical codebook 33 gives a plurality of statistical code vectors Vsk as candidates in time sequence or simultaneously. Thus, from the adder 37, the excitation code vector V to which these vectors Vd and Vsk are added
k (k = 1 to m) is output.

The subsequent circuit processing for the m excitation code vectors Vk thus obtained is the same as the processing for detecting the optimum adaptive code vector Va, and therefore detailed description thereof is omitted. .. Even when the optimum statistical code vector Vs is detected, the gain circuit 40 outputs the same excitation gain g as when the optimum adaptive code vector Va is detected, and the vocal tract analysis circuit 44 also outputs the optimum excitation gain g. The same vocal tract prediction coefficient ai as when detecting the adaptive code vector Va is output.

When the index search circuit 46 detects the optimum statistical code vector Vs, its index I
s is sent to the quantizer 34 and the statistical codebook 33. As a result, the quantizer 34 outputs the total code C.

When the optimum adaptive code vector Va and the statistical code vector Vs are determined in this way, the contents of each part are updated for the processing at the next time.

At this time, both the adaptive codebook 32 and the statistical codebook 33 output the optimum excitation code vectors Va and Vs at the present time. The optimum adaptive code vector Va is the vector V through the thinning circuit 36.
It is given to the adder 37 as d. Thus, adder 3
From 7 the optimum excitation code vector V at the current time is the multiplier 3
9 and the subframe delay circuit 47.

The subframe delay circuit 47 delays this vector V by a subframe and supplies it to the interpolation circuit 38. The interpolation circuit 38 performs an interpolation process on this vector V, and uses an interpolation vector Vn in which the number of components is increased by the above-mentioned X times the number of components of the vector V as the adaptive codebook 32.
Give to. The adaptive codebook 32 updates the contents using this interpolation vector Vn to prepare for the processing at the next time. Since the interpolation processing is performed in this way, the thinning circuit 36 described above is required.

When the optimum excitation code vector V is given from the adder 37 at the present time, the multiplier 39 multiplies the excitation gain g at the present time to obtain the vector Vg, and the vector 39 is added to the adder 41 and the gain control circuit 40. give. Gain control circuit 40
Is the excitation gain g using the subsequent processing using the vector Vg sequence including the current time vector Vg, as described above.
To get

The adder 41 adds the vector Vg and the current-time predicted vector Sp from the synthesis filter 42 to obtain the current-time synthesized speech vector Ss, and supplies it to the synthesis filter 42 and the vocal tract analysis circuit 44. The vocal tract analysis circuit 44 also uses this synthesized speech vector Ss to obtain a vocal tract prediction coefficient ai used in the subsequent processing. The synthesis filter 42 takes in the synthesized speech vector Ss for use in the subsequent predictive synthesis processing.

When the series of processes described above is completed, the process proceeds to the input voice vector S at the next time.

(A-2) Code-Excited Linear Prediction Decoder Applying the Statistical Codebook of the First Embodiment Next, the statistical codebook of the first embodiment is applied and corresponds to the encoder of FIG. The code-excited linear predictive decoder will be described with reference to FIG.

In FIG. 5, the inverse quantizer 50 is provided with the total code C received from the input terminal 51. The dequantizer 50 separates the total code C into an adaptive code vector index Is and a statistical code vector index Ia and supplies them to an adaptive codebook 52 and a statistical codebook 53, respectively.

The adaptive code book 52 outputs the adaptive code vector Va designated by the index Ia, and the thinning circuit 54 thins out the vector Va and outputs the vector Vd to the adder 55. The statistical codebook 53 has the same configuration as the statistical codebook 33 of the encoder, and outputs the statistical code vector Vs designated by the index Is to the adder 55. Thus, the excitation code vector V is obtained, which is given to the multiplier 56 and the subframe delay circuit 57.

The excitation code vector Vp (V) given to the interpolation circuit 58 via the sub-frame delay circuit 57.
Is interpolated by the interpolation circuit 58 and given to the adaptive code book 52, and is used to update the stored adaptive code vector.

The multiplier 56 adds the excitation code vector V to
The excitation gain g given from the gain control circuit 59 is multiplied, and the vector Vg after multiplication is given to the adder 60 and the gain control circuit 59. The gain control circuit 59 also updates the excitation gain g to be given to the multiplier 56 in a predetermined cycle by using the vector Vg after the multiplication.

The output vector Sp from the synthesis filter 61 is also given to the adder 60, and the synthesized speech vector S reproduced by adding the vector Vg and the vector Sp.
s is obtained and given to the hearing correction filter 62, and the hearing correction filter 62 corrects this vector Ss according to the hearing characteristics and outputs it as the final synthesized speech vector Sc from the output terminal 63.

The synthesized voice vector Ss output from the adder 60 is also given to the synthesis filter 61 and the vocal tract analysis circuit 64. The synthesis filter 61 takes in this vector Ss for processing at the next time. The vocal tract analysis circuit 64 obtains the vocal tract prediction coefficient ai, and takes in this vector Ss as information used when updating the vocal tract prediction coefficient ai.

When such a series of processes is completed, the process proceeds to the next reception total code C.

(A-3) Statistical Codebook of First Embodiment Next, details of the statistical codebook 70 (33 or 53) used in the above-described code-excited linear predictive encoder and decoder will be described. explain. Note that FIG. 1 shows the detailed structure of the statistical code book 70.

In FIG. 1, this statistical codebook 70
Is the first and second statistical codebook units 71 and 72
, First and second multipliers 73 and 74, and adder 75
And a weighting coefficient determination unit 76.

The first and second statistical code book sections 71 and 72 store statistical code vectors in consideration of input voices having different characteristics. For example, the first statistical codebook unit 71 stores the statistical code vector considering the average voice of the male, while the second statistical codebook unit 72 stores the statistical code vector considering the average voice of the female. To store. Also, for example, the first statistical codebook unit 71 stores a statistical code vector considering Japanese speech, while the second statistical codebook unit 72 stores a statistical code vector considering English speech. Further, for example, the statistics are calculated such that the change pattern of the vocal tract prediction coefficient a i is different between the case where the first statistical codebook unit 71 is used alone and the case where the second statistical codebook unit 72 is used alone. The code vector is divided into two and stored in the statistical code book units 71 and 72.

When outputting a statistical code vector for a certain index, such as when searching for an index or after determining the optimum index, both the first and second statistical codebook units 71 and 72 are related to that index. The statistical code vectors CBa and CBb are output.

The statistical code vector CBa output from the first statistical codebook section 71 is given to the first multiplier 73, and the weighting coefficient (α) times the weighting coefficient determination section 76 gives to this multiplier 73. It is provided to the adder 75. The statistical code vector CBb output from the second statistical codebook unit 72 is given to the second multiplier 74, and is multiplied by the weighting coefficient (1-α) given to this multiplier 74 from the weighting coefficient determination unit 76. Is given to the adder 75. In this way, the statistical code vectors CBa and CBb output from the first and second statistical codebook units 71 and 72 are weighted and added, and the statistical code vector (= α × CBa + (1-α) × CBb) is added. The statistical code vector obtained by the calculator 75 becomes the output vector of the statistical code book 70.

The weighting coefficient determining unit 76 determines the weighting coefficients α and 1-α from the information about the locally reproduced or reproduced synthetic speech vector output from the adder (41 or 60) described above. Pitch information and vocal tract prediction coefficients ai can be used as the information about the synthesized speech vector. For example, when the statistical code book portions 71 and 72 for men and women are provided, pitch information is extracted and used. Further, for example, when the two statistical codebook units 71 and 72 are provided so that the change pattern of the vocal tract prediction coefficient ai is different from the other code vector unit alone, the vocal tract analysis circuit 44 or 64 described above is used. The change pattern of the vocal tract prediction coefficient a i from the reference pattern is compared with the reference pattern to determine the weighting coefficients α and 1-α.

(A-4) Effects of First Embodiment Therefore, according to the first embodiment described above, the statistical code vectors output from the two statistical codebook units 71 and 72 are used as past synthesized speech vectors. In addition to determining the statistical code vector to be used by weighted addition with the weighting factor determined from the information of, and the information related to the statistical code vector exchanged in the encoder and decoder is only the index, the information to be exchanged is at least In fact, it is possible to obtain a reproduced voice having a quality equivalent to that obtained by using the statistical code vector of the number obtained by multiplying the number of indexes by the changing step of the weighting coefficient. That is, it is possible to realize a device suitable for low bit rate transfer.

(B) Second Embodiment (B-1) Statistical Codebook of Second Embodiment Next, a statistical codebook of the second embodiment will be described. The positions of the statistical codebook of the second embodiment in the encoder and the decoder are the same as those of the first embodiment, for example, and the description thereof will be omitted. FIG. 6 shows the structure of the statistical code book 70A of the second embodiment, and the portions corresponding to those in FIG. 1 are designated by the same reference numerals.

In FIG. 6, when a certain index for searching the optimum index is designated, or when the index determined to be the optimum index is designated again for the next processing, the index conversion unit 7
7 obtains a first index for the first statistical codebook section 71 and a second index for the second statistical codebook section 72 from this index, respectively,
It is given to the corresponding statistical codebook units 71 and 72.

As a result, each statistical codebook section 71, 7
2 outputs statistical code vectors CBa and CBb according to the input index. Subsequent processing is the same as in the case of the first embodiment, so its explanation is omitted.

That is, in the statistical codebook 70A of the second embodiment, the indexes of the first statistical codebook section 71 and the second statistical codebook section 72 are not shared, and the input indexes are those It shows a combination of indexes of the first and second statistical codebook units 71 and 72. That is, it shows the combination of the statistical code vectors in the first and second statistical codebook units 71 and 72. This point is different from the first embodiment. Note that transmission indexes may be prepared for all combinations of the statistical code vectors in the first and second statistical codebook units 71 and 72, but only some of the combinations are prepared. Is also good.

(B-2) Effect of the Second Embodiment Therefore, even if the number of statistical code vectors actually prepared is small according to the second embodiment, it is the same as the case where more statistical code vectors are prepared. It is possible to realize various reproduced voice quality with a small amount of information exchanged between the encoder and the decoder. Therefore, it is possible to realize the one suitable for low bit rate transmission.

(C) Other Embodiments In the above embodiments, the weighting factors α, 1−α are shown to be associated with each other so that they are added to one. Alternatively, a weighting coefficient that is independently determined may be used. That is, the weighting determination unit 76 uses the weighting coefficient β for the first statistical codebook unit 71 and the second statistical codebook unit 72.
It is also possible to determine the weighting coefficient γ for irrelevant.

Further, the number of statistical codebook parts forming the statistical codebook is not limited to two as in the above embodiment, but may be three or more.

Further, the present invention is not limited to the one used in the backward type code excitation linear predictive encoder and decoder as in the above-mentioned embodiment, but the forward type code excitation linear predictive encoding. It may be one used in the decoder and the decoder. In the case of the forward type, since parameters and the like obtained from the input speech vector are transmitted and received by the encoder and the decoder in addition to the index information, the weighting coefficient determination unit 76 of the encoder and the decoder uses this parameter. You may make it determine a weighting coefficient.

[0063]

As described above, according to the present invention, a plurality of statistical code vectors stored corresponding to a certain index are set, and the plurality of statistical code vectors are weighted and added to search for an optimum index. Since the statistic code vector to be used for is output, it is possible to obtain the same quality of reproduced voice as using a larger statistic code vector number with a smaller statistic code vector number,
Moreover, even in such a case, it is possible to realize a statistical codebook in which a small amount of information is exchanged between the encoder and the decoder.

[Brief description of drawings]

FIG. 1 is a block diagram showing a statistical codebook according to a first embodiment.

FIG. 2 is a block diagram showing a conventional code excitation linear predictive encoder.

FIG. 3 is a block diagram showing a conventional code-excited linear predictive decoder.

FIG. 4 is a block diagram showing a code excitation linear predictive encoder to which the statistical codebook of the first embodiment is applied.

FIG. 5 is a block diagram showing a code-excited linear predictive decoder to which the statistical codebook of the first embodiment is applied.

FIG. 6 is a block diagram showing a statistical codebook of the second embodiment.

[Explanation of symbols]

 33, 53, 70, 70A ... Statistical codebook, 71 ... First statistical codebook section, 72 ... Second statistical codebook section, 73, 74 ... Weighted addition multiplier, 75 ... Weighted addition adder, 76 ... Weighting coefficient determination unit, 77 ... Index conversion unit.

Front page continuation (72) Inventor Hiroshi Katsuragawa 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd.

Claims (1)

[Claims] 1. A statistical codebook provided in a code-excited linear predictive encoder or decoder, comprising: a plurality of statistical codebook sections storing different statistical code vectors; A weighted addition unit that forms a statistical code vector by weighted addition of a plurality of output statistical code vectors, and based on the information of the speech vector or the input speech vector already synthesized based on the output statistical code vector And a coefficient determining section that determines a weighting coefficient used by the weighting adding section.