JPH0511799A - Speech coding system - Google Patents

Abstract

(57) [Summary] [Object] An object of the present invention is to speed up a speech encoding process by reducing the amount of calculation at the time of error evaluation while minimizing the deterioration of the quality of synthesized speech. . [Configuration] An excitation vector 102 is selected from the codebook storage means 101. The voice synthesis processing means 103
Speech synthesis processing based on linear prediction analysis is performed using the excitation vector 102 as an input. The decimation unit 106 decimates the element values of the input speech vector 105 and the element values of the synthesized speech vector 104 at regular intervals. The error evaluation means 109 evaluates the squared error for the thinned-out input speech vector 107 and the thinned-out synthesized speech vector 108. The above process is repeated for each excitation vector 102, and the input voice vector 105 is encoded based on the parameter that has generated the synthesized voice vector with the minimum squared error (the maximum correlation value).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coding system for performing information compression transmission of a voice signal, and more specifically,
The present invention relates to an error evaluation method between an input speech vector and a synthetic speech vector in a speech coding method using a code driven LPC coding (CELP) method or the like.

[0002]

2. Description of the Related Art Since voice is an important means in human communication, voice information plays a very important role in society. In the modern age of OA, voice communication by telephone along with numbers and letters is an indispensable communication means. In particular, in recent years, with the remarkable improvement in performance and miniaturization of hardware supported by digital signal processing technology and LSI technology, expansion of digital communication paths and IS
Digital integration of various services using DN (Service Integrated Digital Network) is rapidly progressing in Japan and other countries.

In digital communication, the basic measure is the amount of information transmission (bit rate) per unit time of the line used.
Therefore, in order to efficiently and economically use the network and meet the communication demand, it is necessary to develop communication information compression technology immediately. Given that voice as a means of communication by humans is required to have functions such as promptness / indicativeness / alertness, and is often used for communication such as office communication / activity instruction in society, in particular, In digital mobile radio systems, in-house communication systems, and the like, there is a demand for a voice encoding method capable of realizing highly efficient information compression of voice signals at low and medium bit rates of 4 to 16 Kbits / sec.

As a typical high-efficiency speech coding method, there is an analysis-synthesis coding method in which a speech signal is analyzed based on a speech generation model, parameters are extracted, and coding is performed. In order to perform significant information compression of voice signals, it is necessary to efficiently extract voice information from voice signals, and for that purpose, it is essential to introduce a voice generation model. As a typical and practical analysis / synthesis method based on the speech generation model, there is an analysis / synthesis method based on the linear prediction model.
It is called a (linear predictive coding) method.

This method utilizes the fact that there is a high correlation between the sample values of the speech waveform, and the present signal x _i is a predicted value that is a linear combination of the past K (about 10) samples, and at that time. Is to be expressed as the sum of the error signals of

[0006]

[Equation 1]

The expression can be reduced to: Here, a _j is a linear prediction coefficient (LPC coefficient), e _i is called a prediction error signal, and the LPC coefficient is obtained under the condition that the mean square error of the prediction error signal e _{i in} a certain section is minimized. This is called linear prediction analysis (LPC analysis). Performing LPC analysis on the original speech signal (the speech signal to be encoded) is equivalent to assuming a speech production system based on an all-pole model, the poles of which are mainly in the frequency characteristics of the vocal tract of humans. It corresponds to the peak of the spectrum called the formant. The LPC coefficient is a system parameter when the original speech signal is approximated by a speech generation model by an all-pole model in a certain section (for example, a section of several tens of milliseconds) in which the frequency characteristic of the original speech signal is considered to be stationary. And L for each of the above sections
High-efficiency coding of voice can be realized by calculating LPC coefficients by PC analysis and coding them.

Here, in the above equation 1, the first right side
The term represents the operation of the speech synthesis filter using K LPC coefficients a _j as filter coefficients, and the prediction error signal e _i of the second term on the right side is a system for obtaining the output x _i of the speech synthesis filter. Can be thought of as the input to, ie the sound source.
That is, in order to synthesize speech in the LPC system, in addition to the LPC coefficient for constructing the speech synthesis filter,
A sound source signal that is an input to the speech synthesis filter is required. Therefore, when speech is encoded based on the LPC method, it is necessary to encode the LPC coefficient and efficiently encode the prediction error signal e _i .

Here, the above equation 1 is

[0010]

[Equation 2]

## EQU1 ## From which the prediction error signal e _i can be obtained by passing the original speech signal x _i through an inverse filter having the inverse characteristics of the linear system modeled by LPC analysis. This prediction error signal is also called the residual signal.

FIG. 7 is a diagram showing an example of the time domain waveform of the residual signal in comparison with the original voice signal. If the prediction of the speech signal is perfectly performed in the LPC analysis, the residual signal becomes a completely random signal having a flat power spectrum with a small value. However, in reality, particularly in a voiced sound portion that occupies most of the speech, the speech signal is not completely predicted by LPC analysis, and the residual signal has a time series of periodic pulses as shown in FIG. appear. Also, a noise time series appears between the pulse time series. On the other hand, in the unvoiced part, the residual signal is a noise time series that is almost white in frequency.

As described above, the residual signal includes a pulse time-series component in the voiced sound portion from a macro perspective, and a noise time-series component from a micro perspective. It is known that the pulse time series is synchronized with the pulse train airflow generated by the vibration of the vocal cords of the human throat, and the repetition cycle of the pulse is called the pitch cycle. Also,
The noise time series can be approximately represented by any of typical noise time series candidates statistically classified into a plurality of types for every fixed sample (for example, 64 samples at 8 kHz sampling).

Utilizing such a fact, it is possible to efficiently and faithfully encode a residual signal in a voice transmission system of low to medium bit rate of 4 to 16 kbits / sec applicable to mobile radio or the like. As a voice coding method that can be performed, there is a code driven LPC coding (CELP, hereinafter the same) method.

In the CELP system, the following encoding processing is performed in the transmitting device. That is, first, the LPC coefficient mainly corresponding to the vocal tract characteristic is calculated by LPC analysis for each part (this is called a frame) in which the input audio signal to be transmitted is divided into predetermined samples, and the LPC coefficient is calculated. Is encoded as vocal tract information corresponding to the current frame and transmitted to the receiving device. Next, among the components representing the residual signal obtained by removing the components represented by the LPC coefficient from the input speech signal, the pulse time series components and the remaining noise time series components are encoded separately, It is transmitted to the receiving device on the decoding side. In the receiving device,
The noise time series component is decoded and the pulse time series component is decoded based on each code data relating to the residual signal transmitted from the transmitter, and the residual signal is decoded by adding both components. . Then, the LPC coefficient transmitted from the transmission device constitutes an LPC synthesis filter, and the above-mentioned decoded residual signal is input to the LPC synthesis filter to synthesize the output audio signal.

As described above, in the CELP method, the component corresponding to the vocal tract characteristic of the voice is encoded as the LPC coefficient, and the residual signal corresponding to the source characteristic of the voice is composed of the pulse time series component and the noise time series. By separating and encoding the components, it is possible to encode the voice with high efficiency and high quality.

Here, a codebook storing pulse time series candidates and a codebook storing noise time series candidates are prepared for both the transmitter and the receiver. The transmitter and the receiver prepare the same pulse time series codebook and the same noise time series codebook, respectively.

Then, in the transmitting apparatus, a pulse time series that optimally approximates the frame-by-frame residual signal is searched from candidates in the pulse time series codebook, and the position of the pulse time series on the codebook is searched. Is transmitted to the receiving device. At this time, the optimum gain for adapting the optimum pulse time series amplitude to the amplitude of the residual signal is also calculated and encoded and transmitted to the receiving device. Next, the component obtained by multiplying the optimum pulse time series by the optimum gain is removed from the residual signal, and a residual error signal in frame units is obtained. Then, the noise time series that optimally approximates the residual error signal of this frame unit is searched from among the candidates in the code book of the noise time series in the same manner as in the case of searching for the pulse time series candidates. The index is transmitted to the receiving device. Further, the optimum gain for adapting the optimum noise time series amplitude to the amplitude of the residual error signal is also calculated and encoded and transmitted to the receiving device.

Corresponding to the above operation on the transmitting device side, the receiving device refers to the noise time series codebook with the index of the noise time series sent from the transmitting device, and reads the corresponding noise time series. Be done. Then, the residual error signal is decoded by multiplying the noise time series by the optimum gain sent from the transmitter. Then, the codebook of the pulse time series is referred to by the index of the pulse time series sent from the transmitter, and the corresponding pulse time series is read. Then, the pulse time series is multiplied by the optimum gain sent from the transmitter, and the above-mentioned residual error signal is added to the pulse time series signal obtained as a result, thereby decoding the residual signal. It

Here, in the transmitter, the optimum pulse time series and the optimum gain for approximating the residual signal are searched from the pulse time series codebook, and the residual error signal is obtained from the noise time series codebook. A CELP process for searching for an approximate optimum noise time series and an optimum gain will be described with reference to the principle configuration diagram of FIG. In FIG. 8, reference numeral 801 denotes a part for performing the former process and 802 for a latter process.

First, in order to search the pulse time series that best approximates the residual signal in frame units from the candidates in the pulse time series codebook, simply based on the LPC analysis on the input speech signal. The error between the obtained residual signal of each frame (see Formula 2) and each pulse time series candidate in the codebook may be calculated, and the one that minimizes the error may be selected. However, in practice, the pulse time series that minimizes the human auditory error should be selected, and therefore the error evaluation is performed at the level of the audio signal. Moreover, in that case, the error evaluation is performed not by the audio signal as it is, but by the level of the audio signal whose frequency characteristics are weighted so as to match the human hearing.

That is, in the portion 801 of FIG. 8, first,
The perceptual weighting unit 803 performs weighting on the input voice vector s (input voice signal composed of a plurality of samples) for one frame by weighting its frequency characteristics so as to match human hearing. The attached input speech vector x (weighted input speech signal composed of a plurality of samples) is output.

On the other hand, the pulse vector codebook 807
The (pulse time series codebook) stores a plurality of sets of pulse vectors (pulse time series) each consisting of a plurality of samples. Here, one set of pulse vectors is p (m). m is the pulse vector codebook 80
The index of a plurality of sets of pulse vectors on 7 is represented, and if the number of sets is α, then 1 ≦ m ≦ α.

Now, the error evaluation unit 804 instructs the reading unit 806 to read the first set of pulse vectors with m = 1. As a result, the first set of pulse vectors (pulse time series) p (1) read from the pulse vector codebook 807 (pulse time series codebook) by the reading unit 806 is input to the weighted short-term prediction unit 805. To be done.

The weighted short-term predictor 805 uses the LPC coefficient a.
By performing the processing of the LPC synthesis filter composed of _j (1 ≦ j ≦ K) in consideration of auditory weighting,
A synthetic speech vector y (1) (a synthetic speech signal composed of a plurality of samples) corresponding to the first set of pulse vectors p (1) is calculated. The LPC coefficient a _j (1 ≦ j ≦ K) is obtained by performing the Kth-order LPC analysis process (not shown) on the input speech vector s described above.

The error evaluator 804 calculates an error between the weighted input voice vector x and the synthesized voice vector y (1) corresponding to the above-mentioned first set of pulse vectors p (1). Then, when the error evaluation unit 804 finishes the error evaluation of the first set, the value of m is updated to 2 and the reading unit 806 instructs the pulse vector codebook 807 to set the pulse of the next set corresponding to m = 2. Read the vector p (2).

In this way, the error evaluator 804 uses m
Of the pulse vector p that is sequentially read from the pulse vector codebook 807 while updating the value of 1 from α to α.
For the synthesized speech vector y (m) corresponding to (m), an error between the synthesized speech vector y (m) and the same weighted input speech vector x found by the perceptual weighting unit 803 is calculated, and all pulse vectors p
When the error evaluation for (m) (1 ≦ m ≦ α) is completed, the value m _{opt of the} index m corresponding to the pulse vector having the smallest error and the optimum gain g are output.

As described above, the voice signal (voice vector)
The method of performing error evaluation while synthesizing is generally called an analysis-by-synthesis method (AbS method). With the AbS method, it is not necessary to directly obtain the residual signal from the input speech signal, but it is necessary to execute speech synthesis processing each time the codebook is searched.

Here, the error evaluation method in consideration of the gain in the error evaluation unit 804 will be described. First, the error power Ed between the weighted input speech vector x and the synthesized speech vector y (m) is expressed by the following equation.

[0030]

[Equation 3]

Here, the synthesized speech vector y (m) is calculated by the weighted short-term prediction unit 805.

[0032]

[Equation 4]

Are combined by the following filter calculation. Each subscript i of the weighted input speech vector x and the synthesized speech vector y (m) represents the number of each element of one frame (one set) of speech samples consisting of N samples, and 1 ≦ i ≤
N.

The gain g that minimizes the error power Ed of the above equation 3 is obtained by differentiating the equation 3 with g and setting it to zero. That is, it is as follows.

[0035]

[Equation 5]

The error power Ed at this time is given by the following equation by substituting the equation 5 into the equation 3.

[0037]

[Equation 6]

The first term on the right side of the equation (6) represents the power of the weighted input speech vector x, and each pulse vector p
By changing the set of (m), the synthesized speech vector y
It is constant even if the set of (m) is changed. Therefore, the set of pulse vectors p (m) that minimizes the error power Ed is the second term on the right side of Equation 6, that is,

[0039]

[Equation 7]

It can be obtained as a set of pulse vectors p (m) that maximizes. The expression (7) is an operation for obtaining a kind of correlation between the weighted input speech vector x and the synthesized speech vector y (m). Therefore, obtaining the pulse vector p (m) that maximizes the equation (7) is almost equivalent to obtaining the one having the maximum correlation between the two voice vectors.

Based on the above-described principle of error evaluation, the error evaluation unit 804 of FIG. 8 calculates the equation 7 for the synthetic speech vector y (m) corresponding to each set of pulse vectors p (m), and The value of the index m corresponding to the set in which the value A finally becomes the largest is output as m _opt . That is, number 7
The optimal pulse and synthetic speech vectors that maximize the value of the equation are p (m _opt ) and y (m _opt ).

Then, the error evaluation section 804 outputs the optimum gain g by calculating the expression 5 using the optimum synthesized speech vector y (m _opt ). As a result of the above processing, the above-mentioned optimum synthesized speech vector y (m _opt ) is output from the error evaluation unit 804 to the multiplication unit 808. Then, the multiplication unit 80
8, y (m _opt ) is multiplied by the optimum gain g, and the subtraction unit 809 further subtracts the above-mentioned multiplication result from the weighted input speech vector x _i to obtain the weighted input speech error vector ex. .

Next, even when a noise time series that optimally approximates the residual error signal in frame units is searched from the candidates in the noise time series codebook, at the level of the perceptually weighted speech signal. Error evaluation is performed.

That is, in the portion 802 of FIG. 8, the noise vector codebook 813 (noise time series codebook) stores a plurality of sets of noise vectors (noise time series) each consisting of a plurality of samples. Where 1
Let the noise vector of the set be c (n). n represents an index of a plurality of sets of noise vectors in the noise vector codebook 813, where 1 ≦ n ≦ β where β is the number of sets.

The control operation by the error evaluator 810 is almost the same as that of the error evaluator 804 described above. That is, the error evaluation unit 810 updates the value of n from 1 to β, and synthesizes the weighted short-term prediction unit 811 corresponding to the noise vector c (n) sequentially read from the noise vector codebook 813. For the speech error vector ey (n), the error with the weighted input speech error vector ex is calculated, and at the time when the error evaluation for all the noise vectors c (n) (1 ≦ n ≦ β) is completed, the error becomes The value n _{opt of the} index n corresponding to the smallest noise vector and the optimum gain b are output.

The error evaluation including the gain in the error evaluation unit 810 is performed based on the following Expressions 8 to 12 corresponding to the above Expressions 3 to 7.

[0047]

[Equation 8]

[0048]

[Equation 9]

[0049]

[Equation 10]

[0050]

[Equation 11]

[0051]

[Equation 12]

That is, the error evaluator 810 includes the read unit 81.
2, the weighted short-term predictor 811 corresponding to each set of noise vectors c (n) read from the noise vector codebook 813 according to the equation (9) synthesizes the speech error vector ey (n) with the equation (12). The formula is calculated, and the value of the index n corresponding to the set whose value A finally becomes the largest is output as n _opt . Furthermore, the error evaluation unit 81
0 outputs the optimum gain b by calculating the expression (10) using the optimum synthesized speech error vector ey (n _opt ).

Corresponding to the operation on the transmitting device side as described above, at the receiving device, the noise vector codebook in the receiving device having exactly the same configuration as 813 in FIG. 8 is obtained by the index n _opt sent from the transmitting device. The optimum noise vector c (n _opt ) corresponding to this is read out. Then, the residual error signal is decoded by being multiplied by the optimum gain b sent from the transmitter. Then, with the index m _opt sent from the transmission device, 80 in FIG.
Reference is made to the pulse vector codebook in the receiver having exactly the same configuration as in No. 7, and the corresponding optimum pulse vector p (m
_opt ) is read. Then, it is multiplied by the optimum gain g sent from the transmitter, and the above-mentioned residual error signal is added to the pulse vector obtained as a result, whereby the residual signal is decoded.

[0054]

In the conventional example of the CELP method based on the AbS method described above, it is necessary to execute the speech synthesis processing every time the codebook is searched, so that the optimum index m To obtain _opt , n _opt and optimum gains g, b requires a very large amount of calculation.

Particularly, in order to obtain an excellent synthesized speech quality, the size β of the noise vector codebook 813 shown in FIG. 8 is used.
Needs to be increased. This is for the following reason. That is, there are not so many types of pulse vectors to which the residual signal should be approximated, whereas the residual error signal obtained by removing the pulse vector components from the residual signal has a frequency characteristic close to white and As a vector pattern, a great many kinds of patterns can appear. Therefore, many patterns of noise vectors are required to accurately approximate such a residual error signal.

As described above, as a result of requiring the large-sized noise vector codebook 813, the number of times the error evaluator 810 of FIG. 8 calculates the equation 12 increases, resulting in a huge amount of calculation as a whole. There is a problem that it ends up.

The above-mentioned problems are not limited to the CELP method, and for example, in the multi-pulse coding method which is one of the coding methods of the residual signal, it is synthesized with the input voice based on the AbS method. The same occurs in the process of repeating the evaluation of the square error of the voice and the like.

It is an object of the present invention to speed up the speech coding process by reducing the amount of calculation at the time of error evaluation while minimizing the deterioration of the quality of synthesized speech.

[0059]

FIG. 1 is a block diagram of the present invention. The present invention is premised on a speech encoding method in which an input speech vector is encoded while performing a speech synthesis process and repeating a squared error evaluation process between the synthesized speech vector obtained thereby and the input speech vector. More specifically, for example, a speech coding method based on a code driven linear prediction method is applied. In this method, first, the input speech vector 105 is subjected to linear prediction analysis and its linear prediction coefficient is encoded. Along with this processing, an excitation vector 102 is selected from the codebook storage means 101, a speech synthesis processing is executed in the speech synthesis processing means 103 based on the excitation vector and the above-mentioned linear prediction coefficient, and a synthetic speech vector 104 obtained thereby. And input voice vector 1
The vector quantization of the residual signal of the input speech vector 105 is performed while the evaluation process of the squared error with 05 is repeated. In addition, a multi-pulse coding method, which is one of the coding methods of the residual signal, can be adopted as the speech synthesis method.

First, the thinning-out means 10 for thinning out the element values of the input speech vector 105 and the element values of the synthesized speech vector 104 at regular intervals.
Have six.

Next, an error evaluation for executing a square error evaluation process on the thinned-out input voice vector 107 and the thinned-out synthesized voice vector 108 obtained by thinning out the respective element values of the input voice vector 105 and the synthesized voice vector 104 by the thinning-out means 106. Having means 109. The means calculates the correlation value between the thinned-out input speech vector 107 and the thinned-out synthesized speech vector 108, for example.

By repeating the above-described square error evaluation processing, the input voice vector 105 is encoded based on the parameter that has generated the synthesized voice vector having the minimum square error (the maximum correlation value).

In the above-mentioned configuration of the present invention, the squared error evaluation process may be performed after weighting process based on the auditory characteristic. In particular,
Perceptual weighting is performed on the input speech vector 105, and the perceptually weighted synthetic speech vector 104 is generated by the speech synthesis processing unit 103 performing speech synthesis processing in consideration of the perceptual weighting. Then, thinning processing by the thinning means 106 and error evaluation by the error evaluation means 109 are executed for these two vectors.

[0064]

The input voice vector 10 is generated while repeating the evaluation process of the squared error between the synthesized voice vector 104 and the input voice vector 105 obtained by performing the voice synthesis process.
In the speech coding method for performing the coding of No. 5, the multiplication operation between the respective elements of the synthesized speech vector 104 and the input speech vector 105 is required in the evaluation process of the squared error once. And, for example, in a code driven linear prediction scheme,
If the number of types of the excitation vector 102 stored in the codebook storage means 101 increases in order to improve the sound quality of the decoded speech, the square error evaluation process must be performed for all the excitation vectors, so that a huge calculation is performed as a whole. You need the amount.

Therefore, in the present invention, the speech signal is obtained by thinning out the element values of the input speech vector 105 and the synthesized speech vector 104 by the thinning-out means 106 by utilizing the fact that the proximity correlation is high between adjacent several samples. The error evaluation means 109 performs error evaluation on the thinned-out input speech vector 107 and the thinned-out synthesized speech vector 108.

By this means, it is possible to reduce the amount of calculation for error evaluation to a fraction, while minimizing the deterioration of the quality of synthesized speech on the decoding device side, and speed up the speech coding device. Therefore, the circuit scale can be reduced.

[0067]

Embodiments of the present invention will now be described in detail with reference to the drawings. The present invention is roughly divided into an LPC encoding unit in FIG. 2 and a residual encoding unit in FIG.

In the embodiment of the present invention, FIG.
The configuration and operation of the error evaluation unit 310 are characterized. < Description of LPC Encoding Unit > FIG. 2 is a block diagram of the LPC encoding unit.

The input voice signal sampled at the sampling frequency of 8 kHz is held in the buffer unit 201 as the input voice vector s, for example, one frame = 320 samples at a time. Then, the encoding process on the input speech vector s of 320 samples is executed in real time during 320 sampling periods.

The LPC analysis section 202 performs an LPC analysis on the above-mentioned input speech vector s of 320 samples and outputs a set of LPC coefficients a _j (1 ≦ j ≦ K).
As a concrete method of LPC analysis, for example, PARCOR
An analysis method, an LSP analysis method, or the like can be applied, and the order K of the LPC analysis is, for example, 10th order.

A set of 10th-order LPC coefficients a _j obtained by the LPC analysis section 202 for each 320 samples of the input speech vector s is encoded by the encoding section 203 and then transmitted to a speech decoding device (not shown). To be done.

Also, this encoded LPC coefficient a
_j is decoded in the decoding unit 204 by the same processing as the decoding processing in the speech decoding apparatus, and is used in the processing in the residual coding unit described later. In this way, the decrypted LPC
The coefficient a _j is used because the LPC coefficient a _j is included in the residual signal by including a quantization error that occurs when the LPC coefficient a _j is encoded, and thus the LPC coefficient a _j is encoded.
This is because the quality deterioration of the voice at the time of decoding, which occurs due to the encoding of the coefficient a _j , is minimized. < Description of Residual Encoding Unit > Next, FIG. 3 is a configuration diagram of the residual encoding unit.

In FIG. 3, the parts denoted by reference numerals 301 to 313 are arranged in the order of the numbers 80 in the principle configuration of the CELP system of FIG. 8 described in the section "Prior Art".
It corresponds to each part of 1 to 813. In FIG. 3, the basic configuration is the same as the principle configuration of the CELP system in FIG. 8, but a configuration for implementing the CELP system as an actual residual encoding device is added.

That is, first, one frame (= 320) output from the buffer unit 201 of the LPC encoding unit of FIG.
The input speech vector s for (samples) is weighted by the perceptual weighting unit 303 to its frequency characteristic so as to match the human hearing, and the weighted input speech vector x of 64 samples obtained therefrom is stored in the buffer unit. Held at 314. Then, the 64 held in the buffer unit 314
For the weighted input speech vector x of the sample, 301
The processing of portions 302 and 302 is continuously executed.

Here, the processing for 1 subframe = 64 samples in each part of the perceptual weighting sections 303, 301 and 302 is continuously executed for 5 subframes, whereby the input speech vector s of 1 frame = 320 samples. Has been executed.

Therefore, from the encoding device to the decoding device (not shown), the LPC encoder 320 of FIG.
The 10th-order LPC coefficient a _j is output at a rate of 1 set for each sample, and the set of the optimum index m _opt and the optimal gain g at the rate of 5 sets for 320 samples is the same from the residual encoding unit of FIG. Then, the set of the optimum index n _opt and the optimum gain b is output at a ratio of 5 sets to 320 samples. More specifically, the description of the residual coding process in the first stage is as follows. First, in the portion 301, the following is performed for each subframe having the weighted input speech vector x of 64 samples held in the buffer unit 314. Such an error evaluation process is executed.

The configuration of the error evaluation unit 304 of FIG. 3 is shown in FIG. The error evaluation unit 304 includes an evaluation value calculation unit 410, a maximum evaluation value storage unit 402, a synthesized voice vector storage unit 403, an index storage unit 404, and an index counter unit 40.
5 and a gain calculator 406. Below, FIG.
Will be described with reference to FIG.

First, the adaptive codebook 307
Is similar to the pulse vector codebook 807 in FIG. 8, in which α sets of pulse vectors p (m) (1 ≦ m ≦
α) is stored. Here, the number of elements (the number of samples) of the pulse vector of each set is 64 samples in the above example, that is, p (m) = p _i (m) (1 ≦ i ≦ 64).
Further, in the present embodiment, the pulse vector p (m) is
It is updated by the updating unit 327 described later.

Now, the evaluation value calculation unit 4 in the error evaluation unit 304
01 sets the initial value 1 to the index counter unit 405, and the index counter unit 405 reads out the read unit 3
For 06, m = 1 is output. As a result, the reading unit 306 reads the first set of pulse vectors p (1) from the adaptive codebook 307 and outputs it to the weighted short-term prediction unit 305.

The weighted short-term predictor 305 uses the LPC coefficient a.
By executing the processing of the LPC synthesis filter based on the above-described equation 4 configured by _j (1 ≦ j ≦ K) in consideration of auditory weighting, the first set of pulse vectors p
The synthesized speech vector y (1) corresponding to (1) is calculated. This LPC coefficient a _j (1 ≦ j ≦ K) is obtained by the encoding unit 2 in FIG.
It is supplied from 03.

Evaluation value calculation unit 401 in error evaluation unit 304
Is a synthesized speech vector y (1) (= y _i (1), 1 ≦ i ≦ 64) corresponding to the above-described first set of pulse vectors p (1) and one subframe (=) read from the buffer unit 314 (= 64
Sampled weighted input speech vector x (= x _i , 1
≦ i ≦ 64), the evaluation value A of the above-mentioned expression 7 is calculated.

When the evaluation value calculation section 401 has finished the error evaluation of the first set, the evaluation value A and the first set of synthesized speech vectors y
(m) and the index m = 1 are stored in the maximum evaluation value storage unit 402, the synthesized voice vector storage unit 403, and the index storage unit 404, respectively.

Subsequently, the index counter unit 405
Increments the value of the index m, which is a counter value, from 1 to 2 based on an instruction from the evaluation value calculation unit 401. As a result, the reading unit 306 determines that the second set of pulse vectors p from the adaptive codebook 307 is received.
(2) is read and output to the weighted short-term prediction unit 305.

The evaluation value calculator 401 performs the following control processing while updating the value of m from 1 to α in the index counter 405 as described above. That is,
The evaluation value calculation unit 401 uses the adaptive code book 30.
7, the synthesized voice vector y (m) (= y _i (m), 1 ≦ i ≦ 64) corresponding to the pulse vector p (m) and the weighting for one subframe output from the buffer unit 314. With respect to the input voice vector x (= x _i , 1 ≦ i ≦ 64), the evaluation value A of the above-mentioned expression 7 is calculated. Then, the evaluation value calculation unit 401 compares the evaluation value A with the evaluation value stored in the maximum evaluation value storage unit 402, and if the evaluation value A calculated this time is larger, the maximum evaluation value storage unit 402. Is replaced with the evaluation value A of this time, and the synthesized voice vector storage unit 402 and the index storage unit 40 are replaced.
Each of the contents of 4 is the synthesized voice vector y (m) of this time.
And the value of index m this time. If the evaluation value A calculated this time is equal to or smaller than the evaluation value stored in the maximum evaluation value storage unit 402, the above replacements are not performed.

In this way, all index values m
At the time when the error evaluation for (1 ≦ m ≦ α) is completed, the index storage unit 404 stores the index value m = m _{opt of} the pulse vector that maximizes the evaluation value A of the equation (7).
Is stored, and the synthesized voice vector storage unit 40
In 3, the optimum synthesized speech vector y (m _opt ) at that time is stored.

After that, the evaluation value calculation unit 401 outputs the weighted input speech vector x output from the buffer unit 314.
(= X _i , 1 ≦ i ≦ 64) and the optimum synthesized speech vector y (m _opt ) read from the synthesized speech vector storage unit 403.
(= Y _i (m _opt ), 1 ≦ i ≦ 64) is calculated by the gain calculation unit 406.
Output to. The gain calculation unit 406 outputs the optimum gain g by calculating the above-mentioned equation 5 based on the above two sets of vectors. Then, the optimum gain g is encoded by the encoding unit 318.

As described above, the optimum index m _opt from the evaluation value calculation unit 401 and the encoding unit 318
The coded optimum gains g are sent to the speech decoding apparatus not shown in the figure.

The evaluation value calculation unit 401 uses the optimum index m _opt read from the index storage unit 404.
Is output to the reading unit 306. The reading unit 306 reads the optimum pulse vector p (m corresponding to the optimum index m _opt.
opt ) (= p _i (m _opt ), 1 ≦ i ≦ 64) is read from the adaptive codebook 307, and the buffer unit 322 is read.
Output to. The data in the buffer unit 322 is used for the update processing of the adaptive code book 307 described later.

Following the above processing, the evaluation value calculation unit 401
Is the optimum synthesized speech vector y (m _opt ) (= y _i (m _opt ), 1 ≦ i ≦ read from the synthesized speech vector storage unit 403.
64) is output to the multiplication unit 308 of FIG. Further, the optimum gain g encoded by the encoding unit 318 is input to the multiplication unit 308 after being decoded by the decoding unit 319 by the same decoding process as the decoding process in a speech decoding device (not shown). The reason why the decoded optimum gain is used in this way is to remove the influence of the quantization error due to the coding, as described in the decoding unit 204. Then, the multiplication unit 308 multiplies the optimum synthesized speech vector y (m _opt ) by the value obtained by decoding the optimum gain g by the decoding unit 319, and further, the subtraction unit 309 outputs the above multiplication result from the buffer unit 314. Weighted input speech vector x (= x _i , 1 ≦ i
≦ 64), the weighted input speech error vector ex (= ex _i , 1 ≦ i for one subframe is obtained.
≦ 64) is obtained. Description of the residual coding process in the second stage Next, in the portion 302 of FIG.
The following error evaluation process is executed for each sub-frame of the 64-sample weighted input speech error vector ex held in.

The structure of the error evaluation unit 310 of FIG. 3 is shown in FIG. The error evaluation unit 310 corresponds to the configuration of the error evaluation unit 304 in FIG.
01, the maximum evaluation value storage unit 502, the synthesized speech vector storage unit 503, the index storage unit 504, the index counter unit 505, and the gain calculation unit 506, but the thinning units 507 and 508 are most included in the present embodiment. This is a great feature. This will be described below with reference to FIGS. 3 and 5.

First, the stochastic codebook 3
13 is similar to the noise vector codebook 813 of FIG. 8, in which β sets of noise vectors c (n) (1 ≦ n ≦
β) is stored. Here, the number of elements (number of samples) of the noise vector of each set is 64 samples in this embodiment,
That is, _{c (n) = c i (} n) (1 ≦ i ≦ 64).

Then, the evaluation value calculation unit 501 operates at 4 in FIG.
Similar to 01, the following control process is performed while updating the value of n from 1 to β in the index counter unit 505. That is, the evaluation value calculation unit 501 corresponds to the set of each noise vector c (n) read from the stochastic codebook 313 by the reading unit 512, and is synthesized by the weighted short-term prediction unit 311 based on the equation (9) and buffered. The synthesized speech error vector ey (n) (= ey _i (n), 1 ≦ i ≦ 64) input via the unit 317 and the weighted input speech error for one subframe output from the buffer unit 316. An error evaluation value is calculated for the vector ex (= ex _i , 1 ≦ i ≦ 64). Then, the evaluation value calculation unit 501 compares the evaluation value with the evaluation value stored in the maximum evaluation value storage unit 502. If the evaluation value calculated this time is larger, the evaluation value calculation unit 501 stores the maximum evaluation value storage unit 502. The contents are replaced with the evaluation value of this time, and the contents of the synthetic speech error vector storage unit 503 and the index storage unit 504 are replaced by
The values of the synthetic speech error vector ey (n) of this time and the index n of this time are respectively replaced. When the evaluation value calculated this time is equal to or smaller than the evaluation value stored in the maximum evaluation value storage unit 502, the above replacements are not performed.

Here, the arithmetic expression of the evaluation value in the evaluation value arithmetic unit 501 is determined based on Equations 3 to 6 which are the arithmetic expressions of the evaluation value in the evaluation value arithmetic unit 401 of FIG. Corresponding to the above, it is determined based on the above-described equations (8) to (11). Then, the arithmetic expression of the evaluation value that is basically derived based on these expressions is the expression 12 used in the conventional example. A major feature of the present embodiment is that the equation (12) is modified into the following equation (13).

[0094]

[Equation 13]

That is, in the conventional example, the weighted input speech error vector ex (= e
x _i , 1 ≤ i ≤ 64) and one set of synthesized speech error vectors ey (n) (= ey _i (n), 1 ≤ i ≤ 64) for each element value (signal sample value) The evaluation value A was calculated by using. On the other hand, in the present embodiment, the evaluation value A is calculated while the element values of each vector are thinned out at a constant interval M, as in Expression 13. In general, it is well known that the audio signal and the synthetic audio signal have a proximity correlation of the order of 8 to 10. Therefore, even if the evaluation value A is calculated while performing the thinning-out process as in Expression 13, the evaluation value A is calculated as compared with the case where the evaluation value A is calculated without performing the thinning-out process as in Expression 12. It is expected that it will not change significantly. Here, “N (step M)” of the equation 13
Indicates that the number i indicating the element value of the vector is incremented from 1 to the value N by the value M.

In the present embodiment, in the thinning-out operation based on the expression (13) up to M = 4, the S / N ratio of the decoded voice is almost equal to that in the case where the operation of the expression (12) is performed. It has been experimentally confirmed that (or segmental S / N) can be obtained.

Here, comparing the calculation amounts when the evaluation value A is calculated by each of the equations 12 and 13, the result is as shown in FIG. From this, for example, when the number of samples in one subframe is N = 64 and the thinning interval is M = 4, it is possible to calculate the evaluation value A by the formula 13 with about 1/4 the number of additions and multiplications. I understand.

Based on the above principle, the error evaluation section 310 shown in FIG. 3 or 4 performs the following arithmetic processing of the evaluation value A. That is, first, the weighted input speech error vector ex for one subframe output from the buffer unit 316
(= Ex _i , 1 ≦ i ≦ 64) is thinned out by the thinning unit 507 at the sample interval M and input to the evaluation value calculation unit 501. In addition, 1 output from the buffer unit 317
A set of synthesized speech error vectors ey (n) (= ey _i (n), 1 ≦
i ≦ 64) is also thinned by the thinning unit 508 at the sample interval M and input to the evaluation value calculation unit 501. These processes correspond to processes in which the number i indicating the element value of the vector is incremented from 1 to the value N by the value M in the equation (13). Then, the evaluation value calculation unit 50
1 executes the remaining arithmetic processing other than the above-mentioned operation in the expression (13).

The synthesized voice error vector storage unit 503
Is not the decimated synthetic speech error vector, but the decimated synthetic speech error vector ey (n) (= ey
_i (n), 1 ≦ i ≦ 64) is stored. This is because the calculation of the optimum gain described later is performed without thinning
It is possible to improve the sound quality of the speech synthesized by the speech decoding device, and the calculation of the optimum gain is 1 for each subframe.
This is because the calculation amount does not increase so much even if the thinning-out process is not performed because it is only executed once.

Based on the calculation processing of the evaluation value A as described above, at the time when the error evaluation for all index values n (1≤n≤β) is completed, the index storage unit 504
Stores the index value n = n _{opt of} the noise vector that maximizes the evaluation value A of Equation 13, and the synthesized speech error vector storage unit 503 stores the optimum synthesized speech error vector ey (n _opt ) (= ey _i (n), 1 ≦ i ≦
64) is stored.

After that, the evaluation value calculation unit 501 reads the weighted input speech error vector ex _i (= ex _i , 1 ≦ i ≦ 64) output from the buffer unit 316 and the synthesized speech vector storage unit 403. The optimum synthesized speech error vector ey (n _opt ) (= ey _i (n _opt ), 1 ≦ i ≦ 64) is output to the gain calculation unit 506. Gain calculation unit 506
Outputs the optimum gain b by calculating the equation 10 similar to the conventional example based on the above two sets of vectors. Then, the optimum gain b is encoded by the encoding unit 320.

As described above, the optimum index n _opt from the evaluation value calculation unit 501 and the coding unit 320
The coded optimum gains b are transmitted from the respective units to a voice decoding device (not shown). Description of Update Process of Adaptive Codebook 307 Lastly, the update process of the adaptive codebook 307 will be described.

The evaluation value calculation unit 501 outputs the optimum index n _opt read from the index storage unit 504 to the reading unit 312 after completing the above-described error evaluation processing for one subframe (= 64 samples). The reading unit 312 uses the optimum noise vector c (n _opt ) (= c _i (n _opt ), 1 ≦ i ≦ 6 corresponding to the optimum index n _opt.
4) is read from the stochastic codebook 313 and output to the multiplication unit 324. On the other hand, the buffer unit 32
From 2 to the multiplication unit 323, the optimum noise vector c (n _opt ) of 64 samples of the optimum pulse vector p (m _opt ) (= p _i (m _opt ), 1 ≦ i ≦ 64) of 64 samples is _output . The corresponding part is output. In the multiplication unit 323, the optimum pulse vector p (m _opt ) is multiplied by the value obtained by decoding the optimum gain g in the decoding unit 319, and in the multiplication unit 324, the optimum noise vector c (n _opt ) is decoded by the decoding unit 321. The multiplied value is multiplied. Here, the reason why the decoded optimum gain is used is to remove the influence of the quantization error due to the coding, as described in the decoding unit 204. Then, these two multiplication results are added by the addition unit 325,
It is stored in the buffer unit 326.

The update data for 64 samples obtained in the buffer unit 326 in this manner expresses an optimum residual signal which will be obtained by a speech decoding device not shown in the figure in the current section of 64 samples. Will be there. Then, the optimum residual signal is written in the adaptive codebook 307 as a new set of pulse vectors by the updating unit 327. At this time, the oldest set of pulse vectors in the codebook is discarded.

As described above, the adaptive codebook 3
In 07, the optimum residual signal set obtained up to the current frame in the processing of 301 in FIG. 3 is stored in the new order, and these data are used for searching the pulse vector in the next subframe. This is because the residual signal has a similar shape between adjacent subframes, and therefore, by using these data in the first stage approximation of the residual signal, a more accurate residual signal This is because encoding can be performed. It should be noted that, in particular, the audio decoding device (not shown) also updates the adaptive codebook in exactly the same manner, so that no contradiction occurs in the data.

In the above description, the embodiment in which the present invention is applied to the CELP system, which is one of the audio coding systems, has been described. However, the present invention is not limited to this, and for example, a residual signal coding system. In the multi-pulse coding method which is one of the above, the same processing can be applied to the processing of repeating the evaluation of the square error of the speech synthesized with the input speech based on the AbS method.

[0107]

According to the present invention, the thinning-out means obtains the thinning-out means by thinning out the respective element values of the input speech vector and the synthesized speech vector by utilizing the fact that the adjacent signal has a high close correlation between several samples. The error evaluation means 109 performs the error evaluation on the input speech vector and the thinned-out synthesized speech vector, so that the deterioration of the quality of the synthesized speech on the decoding device side is minimized, and the calculation amount for the error evaluation is several minutes. It becomes possible to increase the speed of the speech coding apparatus and reduce the circuit scale.

[Brief description of drawings]

FIG. 1 is a block diagram of the present invention.

FIG. 2 is a configuration diagram of an LPC encoding unit in the speech encoding apparatus according to the present invention.

FIG. 3 is a configuration diagram of a residual coding unit in the speech coding apparatus according to the present invention.

FIG. 4 is a configuration diagram of an error evaluation unit 304 in the speech encoding device according to the present invention.

[Fig. 5] Fig. 5 is a configuration diagram of an error evaluation unit 310 in the speech encoding device according to the present invention.

FIG. 6 is a diagram comparing the calculation amounts of Expression 12 and Expression 13.

FIG. 7 is a diagram showing an audio signal and a residual signal.

FIG. 8 is a principle configuration diagram of a CELP method.

[Explanation of symbols]

101 code book storage means 102 excitation vector 103 voice synthesis processing means 104 synthetic speech vector 105 input speech vector 106 thinning means 107 Decimated input voice vector 108 Thinned-out synthetic speech vector 109 Error evaluation means

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Hideaki Kurihara             1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture             Within Fujitsu Limited (72) Inventor Fumio Amano             1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture             Within Fujitsu Limited

Claims (4)

[Claims] 1. A speech coding method for performing coding of an input speech vector while repeating a squared error evaluation process between a synthesized speech vector obtained by the speech synthesis processing and an input speech vector, wherein: Thinning means for thinning out a constant interval for each element value of the vector and the element value of the synthesized speech vector, and obtained by thinning out each element value of the input speech vector and the synthesized speech vector by the thinning means. A speech coding method comprising: an error evaluation unit that executes a squared error evaluation process for a thinned-out input speech vector and a thinned-out synthesized speech vector. 2. An excitation vector (102) is selected from a codebook storage means (101), a speech synthesis process (103) based on linear prediction analysis is performed based on the excitation vector, and a synthesized speech vector ( 104) and the input speech vector (105), while repeating the evaluation processing of the squared error, in a speech coding method based on a code-driven linear prediction method including a process of performing vector quantization of the residual signal of the input speech vector, A thinning-out means (106) for performing thinning-out processing at fixed intervals for each of the element values of the input speech vector (105) and the synthesized speech vector (104), and the input speech vector (105) by the thinning-out means. )
And error evaluation means (109) for executing a squared error evaluation process for the thinned-out input voice vector (107) and the thinned-out synthesized voice vector (108) obtained by thinning out each element value of the synthesized voice vector (104), A speech coding method characterized by having. 3. An excitation vector is selected from the codebook storage means, a speech synthesis process is performed based on the linear prediction analysis based on the excitation vector, and a squared error between a synthesized speech vector and an input speech vector obtained thereby is evaluated. In a speech coding method based on a code-driven linear prediction method, which includes a processing of performing vector quantization of a residual signal of the input speech vector while repeating the processing after weighting processing based on the auditory characteristics, the input speech vector Thinning-out means for performing thinning-out processing at fixed intervals for each of the element values of the above and the synthesized speech vector, and thinning-out obtained by thinning out the respective element values of the input speech vector and the synthesized speech vector by the thinning-out means. Performs squared error evaluation processing on the input speech vector and the thinned-out synthesized speech vector Speech coding, characterized in that it comprises a differential evaluation means. 4. The error evaluation means calculates a correlation value between the thinned-out input speech vector and the thinned-out synthesized speech vector, and obtains a synthesized speech vector having the maximum correlation value as a parameter. The speech coding method according to any one of claims 1 to 3, wherein the input speech vector is coded based on the input speech vector.