JPH11327596A - Voice coding and decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice coding method obtaining a periodic noise code vector cycled by a correct pitch period, without the necessity of transmitting additional information related to pitch cyclization from a coding side to a decoding side and not increasing a code amount. SOLUTION: In a voice coding system respectively inputting an adaptive code vector generated by an adaptive code book 101 and the periodic noise code vector obtained from a noise code book 102 through a pitch periodic part 104 to gain multipliers 105, 106, taking a linear sum with an adder 107, driving a synthetic filter 108 with an obtained drive signal, generating a synthetic voice approximating an input voice and outputting an index specifying the adaptive code vector and the periodic noise code vector from a evaluation part 111, by pitch analyzing the adaptive code vector with a pitch period re-calculation part 103, the pitch period for performing the pitch cyclization of the noise code vector with the pitch periodic part 104 is obtained by calculation.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech encoding / decoding method based on the CELP method or the like, and more particularly, to a technique for pitching noise code vectors for forming a driving signal for driving a synthesis filter.

[0002]

2. Description of the Related Art A method of performing linear prediction analysis on speech and decomposing the speech into linear prediction coefficients representing a spectral envelope and a residual signal is used.
It has been used since ancient times. In recent years, the CELP (Code Excited Linear Prediction (M.P.) method, which has been actively studied as a coding technique for compressing voice or musical sound into a small amount of information and transmitting or storing the same in mobile phones or the Internet.
R. Schroeder and BSAtal, “Code Excited Linear Pre
diction (CELP): High Quality Speech at Very Low Bi
t Rates, “Proc. ICASSP, pp. 937-940, 1985 (Reference 1), and WSKleijin, DJKrasinskl et al.“ Im
proved Speech Qualityand Efficient Vector Quantiza
Option in SELP, “Proc. ICASSP, pp. 155-158, 1988 (Reference 2)) is one of them.

In the CELP system, an input speech signal is divided into a linear prediction coefficient representing phoneme information and a prediction residual signal representing a pitch or the like by linear prediction analysis. By forming a recursive digital filter (synthesis filter) based on the linear prediction coefficients and inputting the prediction residual signal to the synthesis filter, the original input speech signal is restored as a synthesized speech signal.

In order to encode an input speech signal at a low rate, it is necessary to encode these linear prediction coefficients and prediction residual signals with a smaller amount of information. In the CELP method, a signal obtained by encoding a prediction residual signal is generated as a driving signal of a synthesis filter by multiplying two types of signals, an adaptive code vector and a noise code vector, by an appropriate gain and then taking a linear sum. . The method of generating the adaptive code vector is described in detail in Reference 2. The random code vector is usually generated by storing a large number of candidates in a codebook and selecting an optimum one from these. In selecting these code vectors, a synthetic speech signal is generated by passing the code vector through a synthesis filter without directly calculating an error between the code vector and the residual signal, and an error between the synthesized speech signal and the input speech signal is generated. Is a feature of the CELP method. At this time, it is usual that auditory weighting is applied to the error.

[0005] In order to improve the sound quality of synthesized speech, pitch emphasis may be performed by making a pitch period of a noise code vector. As a method of pitch periodizing a noise code vector, a method using a pitch emphasis filter, a method of repeating a part of a noise code vector (PSI conversion) (Miki et al., “CELP coding (PSI conversion with pitch synchronous noise excitation source) -C
ELP) ", Proc. Of the IEICE Spring Conference, S
A-5-5, 1993 (Reference 3)).
In the method using the pitch emphasis filter, a form in which a pitch period component is superimposed on a code vector is used. In the case of PSI, a waveform having a pitch period length is repeated.

[0006] In any of these pitch periodizing methods, the quality is improved by reflecting the pitch component in the noise code vector. Conventionally, the pitch period obtained by the search for the adaptive code vector is used as it is for the pitch period of the noise code vector. By performing such a pitch cycle of the noise code vector, the pitch component of the voice is further emphasized, and a clear synthesized voice is obtained.

In the search for an adaptive code vector,
In some cases, a pitch that is an integral multiple of the basic pitch cycle (referred to as a double pitch) is obtained as the pitch cycle. In the search for the adaptive code vector, a pitch period that minimizes distortion from the target signal including the double pitch is selected. Even if double pitch is selected as the adaptive code vector, it does not pose a fatal problem. This is because the period of the past drive signal constituting the adaptive codebook is T
In this case, even if a waveform having a period of 2T is cut out from this and repeated at 2T, the period of the repeated signal is T.

On the other hand, conventionally, as described above, the noise code vector is pitch-performed with the pitch period obtained by searching for the adaptive code vector. Therefore, when the double pitch is selected in the search for the adaptive codebook, the periodicization of the noise code vector is also performed at the double pitch. However, since the noise code vector originally has no periodicity, in this method, the period of the periodic noise code vector obtained by pitch periodicization is doubled, and it may not be periodic at the correct pitch period. Problems arise.

As a method for alleviating this problem, a plurality of different periods, for example, T
There is also a method in which the pitch period is finally set to a period in which the distortion with respect to the target vector is reduced. However, in this method, it is necessary to transmit, as additional information, information indicating which cycle has been used to perform the pitch period from the encoding side to the decoding side, which causes a new problem that the code amount increases. .

[0010]

As described above, in the conventional pitch period technique, the noise code vector is pitch-performed with the pitch period obtained by searching for the adaptive code vector. When the double pitch is selected in the above, there is a problem that the random code vector is also periodicized at the double pitch, so that the periodicization at the correct pitch period is not performed. However, in the method of finally performing the pitch period in a period in which the distortion with respect to the target vector becomes small, there is a problem that the code amount increases.

The present invention solves such a problem of the conventional pitch periodicity, and can obtain a periodic noise code vector which is periodicized at a correct pitch period. An object of the present invention is to provide a speech encoding / decoding method that does not require transmission of additional information related to pitch periodization and does not increase the code amount.

[0012]

In order to solve the above-mentioned problems, a speech encoding method according to the present invention provides a speech synthesis method which approximates an input speech in frame units by a linear sum of an adaptive code vector and a periodic noise code vector. In the speech coding method for generating an adaptive code vector and an index for specifying a noise code vector, the adaptive code vector is generated based on a drive signal consisting of a past drive signal. 1 and generates a periodic noise code vector with a second pitch period determined independently of the first pitch period with respect to the noise code vector obtained from the noise codebook. It is a basic feature to generate by applying.

Further, the speech decoding method according to the present invention corresponding to this speech encoding method is characterized in that the synthesized speech approximating the input speech per frame is converted into a drive signal comprising a linear sum of an adaptive code vector and a periodic noise code vector. In the speech decoding method according to the present invention, an adaptive code vector is generated from an adaptive codebook including past drive signals based on a first pitch period obtained from an input index, and a periodic noise code vector is generated. The noise code vector is generated by subjecting a random code vector obtained from the random code book to a pitch period at a second pitch period determined independently of the first pitch period based on the input index. .

More specifically, in the speech coding method and the speech decoding method according to the present invention, a second pitch period for converting a noise code vector into a pitch period is determined by analyzing an adaptive code vector (pitch analysis). Ask.

As an analysis method, for example, a correlation value based on the autocorrelation of the adaptive code vector is obtained, and when the correlation value in a pitch cycle shorter than the first pitch cycle is equal to or larger than a predetermined threshold value, the correlation value is calculated. Is given as the second pitch cycle, and when the correlation value at a pitch cycle shorter than the first pitch cycle is less than the threshold value, the first pitch cycle can be taken as the second pitch cycle. .

Further, a correlation value based on the autocorrelation of the adaptive code vector is obtained, and when the correlation value at a pitch period that is an integral fraction of the first pitch period is equal to or greater than a predetermined threshold value, a pitch at which the correlation value is given. The cycle may be a second pitch cycle, and the first pitch cycle may be the second pitch cycle when the correlation value at a pitch cycle that is a fraction of the first pitch cycle is less than the threshold.

As described above, in the speech coding method and the speech decoding method according to the present invention, an adaptive code vector is generated by searching an adaptive codebook for a second pitch period for pitch period of a noise coded vector. Independently of the first pitch cycle used at this time, the pitch is determined based on, for example, the result of the pitch analysis of the adaptive code vector.

Therefore, even when, for example, a double pitch of the basic pitch period is selected as the first pitch period in the search for the adaptive code vector, the pitch period of the noise code vector is changed to the correct pitch period (typically In this case, the pitch period of the random code vector can be accurately changed at the correct period.

Further, if the pitch analysis of the adaptive code vector is performed only in the vicinity of a period that is an integer fraction of the first pitch period obtained in the search of the adaptive codebook, the calculation amount can be further reduced. A second pitch period for the pitch period of the random code vector can be obtained.

Further, according to the present invention, the second pitch period for periodicizing the noise code vector can be obtained by the same algorithm on the decoding side as on the encoding side.
There is no need to transmit extra information on the pitch period of the noise code vector from the encoding side to the decoding side, and there is no problem of an increase in the amount of transmission.

[0021]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a speech encoding system to which a speech encoding method according to an embodiment of the present invention is applied. This speech coding system includes an input terminal 100 to which a speech signal to be coded (input speech signal) is inputted in frame units, an adaptive codebook 101, a noise codebook 102, a pitch period recalculation unit 103, a noise code vector Pitch Periodizing Unit 104 for Pitch Periodization, Gain Multiplier 10
5, 106, an adder 107, a synthesis filter 108, a subtractor 109, an auditory weight filter 110, and an evaluator 111
Consists of

The basic operation of the speech coding system is as follows: an adaptive code vector generated by the adaptive codebook 101;
After passing the periodic noise code vectors obtained from the noise codebook 102 through the pitch periodicizing section 104 through gain multipliers 105 and 106, respectively, the adder 107 calculates the linear sum of the two and drives the synthesis filter 108. By generating a signal and driving the synthesis filter 108 with the driving signal, a synthesized speech signal approximating the input speech signal is generated, and an adaptive code vector index and a noise code for identifying optimal adaptive code vectors and a noise code vector are generated. The vector index is output from the evaluation unit 111.

Next, the processing procedure of the speech coding system according to the present embodiment will be described in detail with reference to FIG. The flowchart of FIG. 2 particularly shows a processing procedure related to the pitch period of the random code vector. In adaptive codebook 101,
A past drive signal for driving the synthesis filter 108 is stored. In order to generate a synthesized speech signal approximating the input speech signal in frame units to the input terminal 100, first, the adaptive codebook 101 is searched. Adaptive codebook 1
In the search for 01, an adaptive code vector is generated by repeating the past drive signal within the determined pitch period. Then, the adaptive code vector is used as the gain multiplier 10
5 and an input to the synthesis filter 107 through the adder 106, an error between the synthesized speech signal obtained by the synthesis filter 107 and the input speech signal is obtained by the subtractor 109, and an error obtained by weighting the error by the auditory weight filter 110 Is evaluated by the evaluation unit 111, and the pitch period T that has generated the adaptive code vector that minimizes this error is selected from the adaptive codebook 101 (step S
1).

Next, the adaptive code vector obtained in the search of adaptive codebook 101 is input to pitch period recalculating section 103, where pitch analysis is performed and pitch period used for pitch period of noise code vector is calculated. K is determined (steps S2 to S6). The procedure for determining the pitch period K will be described later in detail.

In the search for the random codebook 102, the random code vector obtained from the random codebook 102 is pitch-periodized by the pitch period K at the pitch period generator 104 to generate a periodic noise code vector. . The periodic noise code vector is used as the gain multiplier 106
A noise code vector that is convolved by the synthesis filter 108 through the adder 107 and that reduces the perceptually weighted error from the input speech signal in the same manner as in the search of the adaptive codebook 101 is selected.

Finally, the adaptive code vector index and the random code vector for specifying the adaptive code vector and the noise code vector selected in this way are evaluated by the evaluator 111.
Output from The gains to be multiplied by the adaptive code vector and the periodic noise code vector by the gain multipliers 105 and 106 are obtained by searching a gain codebook (not shown), and a gain index for specifying these is output at the same time. .

Here, the feature of the present embodiment is that the pitch period of the noise code vector is changed to the adaptive codebook 101 as in the prior art.
Pitch period T (first pitch period) obtained by search for
Is not performed directly using the pitch period recalculation unit 10
3 is performed by using a pitch period K (second pitch period) determined independently of the pitch period T which is a coding parameter of the current frame by recalculation based on the analysis of the pitch period. By doing so, adaptive codebook 101
Even when the double pitch is selected as the adaptive code vector in the search for, the noise code vector can be periodicized at the basic pitch period, and the sound quality of the synthesized speech can be improved.

In this embodiment, the pitch period recalculating section 103 recalculates the pitch period K for making the pitch period of the noise code vector based on the adaptive code vector. The same pitch period K can be obtained by recalculating. In other words, there is no need to transmit information on the pitch period K for making the pitch period of the noise code vector from the encoding side to the decoding side, and the code amount (bit rate) does not increase.

As a method of pitch-periodizing the noise code vector in the pitch-periodizing unit 104, a method using a pitch filter can be mentioned. Examples of the pitch filter include a recursive filter represented by Expression (1) and a filter of Expression (2).
And the like.

[0030]

(Equation 1)

Here, K 'is an integer part of the pitch period K, and M indicates that the order of the pitch filter is (2M + 1) order. The pitch period K for periodicizing the random code vector takes an integer value or a decimal value depending on the configuration. In the case of an integer value, K = K '. As another method of pitch-periodizing the random code vector, the PSI conversion described in the above reference 3 can be used.

On the other hand, as a method of pitch analysis performed by the pitch period recalculation unit 103, a method using autocorrelation is exemplified. FIGS. 3 and 4 show examples of the relationship between the pitch period and the correlation value based on the autocorrelation. Since the adaptive code vector is originally generated by repeating the same waveform at the period T, the correlation value becomes 1 at an integral multiple of the period T. Therefore, when the period T is the basic pitch period, a peak of the correlation value appears at the period T as shown in FIG.

However, when the period T is twice the basic pitch period (double pitch), the peak of the correlation value still appears at the period T as shown in FIG.
(Basic pitch period) also shows a considerable correlation.
This is because the waveform of the period T is composed of two waveforms having a similar length T / 2, and the autocorrelation value becomes large even at the period T / 2 as shown in FIG. In such a case, it is desirable to set the pitch period K used for the pitch period of the random code vector to be T / 2, which is the basic pitch period.

More specifically, as shown in the flow chart shown in FIG. 2, after the adaptive codebook 101 is searched to generate an adaptive code vector to obtain the pitch period T, the correlation value based on the autocorrelation of the adaptive code vector is obtained. Is calculated (step S2),
It is determined whether the correlation value at T / n (n is an integer of 2 or more) is equal to or greater than a threshold value L (step S3). Here, in step S3, if there is a pitch cycle that gives a peak of a correlation value larger than the threshold value L at a pitch cycle of T / n, K
= T / n (step S4), and if not present, K =
T is set (step S5). Then, the pitch cycle of the noise code vector is performed with the pitch cycle K determined in this way (step S6).

In step S3, when there are a plurality of pitch periods satisfying the condition that the correlation value at T / n is equal to or greater than the threshold value L, in other words, when there are a plurality of n that satisfy this condition, the largest correlation A method such as using a value or using the smallest cycle may be used.

It should be noted that the determination in step S3 is generalized to a determination as to whether or not there is a pitch period T1 that gives a correlation value equal to or greater than the threshold value L with a pitch period shorter than the period T. A method may be adopted in which K = T, and if there is, K = T1. Also in this method, when a plurality of pitches T1 exist (usually, T1
Is n times the pitch), a method such as using the largest correlation value or using the smallest cycle may be used.

(Regarding the Decoding Side) FIG. 6 is a diagram showing a voice decoding system to which the voice decoding method according to the present embodiment is applied, and corresponds to the voice coding system shown in FIG. This speech decoding system includes an input terminal 200 to which an index from the encoding side is input, an adaptive codebook 201, a noise codebook 202, a pitch period recalculation unit 203
And the pitch periodicization unit 204, the gain multipliers 205 and 20
6, an adder 207 and a synthesis filter 208.

The basic operation of the speech decoding system is as follows: an adaptive code vector generated by adaptive code book 201 based on an adaptive code vector index from the encoding side;
The periodic noise code vector obtained by the noise codebook 202 based on the noise code vector index from the encoding side and further obtained through the pitch periodic unit 204 passes through gain multipliers 205 and 206, respectively, and then is added. Vessel 2
In step 07, the synthesis filter 208 is driven by a drive signal obtained by calculating the linear sum of the two, thereby generating a synthesized speech signal that approximates the input speech signal input on the encoding side.

Next, the processing procedure of the speech decoding system according to this embodiment will be described. The adaptive codebook 201 stores past drive signals for driving the synthesis filter 208. An adaptive code vector index and a noise code vector index transmitted from the speech coding system of FIG. 1 are input to an input terminal 200, and based on these indices, an adaptive codebook 201 and a random codebook 202 output an adaptive code vector and a noise codebook 202, respectively. The vector and the random code vector are decoded.

The adaptive code vector from adaptive codebook 201 is input to pitch period recalculating section 203, where the pitch period is recalculated to obtain pitch period K for making the pitch period of the noise code vector. Can be Further, the noise code vector from the noise code butterfly 202 is pitch-periodized by the pitch period K in the pitch periodization unit 204 to become a periodic noise code vector. The adaptive code vector and the periodic noise code vector are respectively assigned to the gain multiplier 20
After the gain is multiplied by 5, 206, the adding section 207
, And becomes a drive signal for the synthesis filter 208.
This drive signal becomes a synthesized speech signal by passing through the synthesis filter 208.

The input signal of the pitch period recalculation unit 203 is an adaptive code vector, and since the adaptive code vector is the same as that on the encoding side, the pitch period recalculation unit 203
In, the same pitch period K as that determined on the encoding side can be obtained by the same algorithm as the encoding side. Therefore, there is no need to transmit information of the pitch period K from the encoding side to the decoding side, and the decoding side always uses the pitch period K corresponding to the basic pitch without requiring any additional information. It is possible to perform the pitch period.

[0042]

As described above, according to the present invention, even when a double pitch is obtained in the search of the adaptive codebook, a periodic noise code vector pitch-performed at a correct pitch period, for example, a basic pitch period. Can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech encoding system to which a speech encoding method according to an embodiment of the present invention has been applied.

FIG. 2 is a flowchart showing a processing procedure related to pitch period in the speech coding system of FIG. 1;

FIG. 3 is a diagram illustrating an example of a relationship between a pitch period and a correlation value based on autocorrelation of an adaptive code vector.

FIG. 4 is a diagram showing another example of the relationship between the pitch period and the correlation value based on the autocorrelation of the adaptive code vector.

FIG. 5 is a diagram showing correlation of adaptive code vectors.

FIG. 6 is a block diagram showing a configuration of a speech decoding system to which the speech decoding method according to the embodiment is applied.

[Explanation of symbols]

 Reference Signs List 101: Adaptive codebook 102: Noise codebook 103: Pitch period recalculation unit 104: Pitch periodization unit 105, 106 ... Gain multiplier 107 ... Adder 108 ... Synthesis filter 109 ... Subtractor 110 ... Auditory weight filter 111 ... Evaluation Unit 201: adaptive codebook 202: noise codebook 203: pitch period recalculation unit 204: pitch period generator 205, 206 ... gain multiplier 207: adder 208: synthesis filter

Claims (8)

[Claims] A synthetic speech approximating an input speech per frame is generated based on a drive signal comprising a linear sum of an adaptive code vector and a periodic noise code vector, and an index for specifying the adaptive code vector and the noise code vector is generated. In the audio coding method to be output, the adaptive code vector is generated based on a first pitch period obtained by searching for an adaptive code book including a past drive signal, and the periodic noise code vector is generated from the noise code book. A speech coding method characterized by generating the noise code vector by subjecting the obtained noise code vector to a pitch period with a second pitch period determined independently of the first pitch period. 2. A speech decoding method for generating a synthesized speech approximating an input speech in a frame unit based on a drive signal comprising a linear sum of an adaptive code vector and a periodic noise code vector, wherein the adaptive code vector is inputted. Based on the first pitch period obtained from the index obtained from the adaptive codebook composed of past drive signals, and converts the periodic noise code vector into a noise code vector obtained from the noise codebook based on the input index. On the other hand, a speech decoding method characterized by generating by performing pitch periodization at a second pitch period determined independently of the first pitch period. 3. A synthetic speech approximating an input speech in a frame unit is generated based on a drive signal composed of a linear sum of an adaptive code vector and a periodic noise code vector, and the adaptive code vector and the noise code vector are specified. A speech encoding method for outputting an index, wherein the adaptive code vector is generated based on a first pitch period obtained by searching for an adaptive codebook including past drive signals, and the periodic noise code vector is generated by a noise code. A speech encoding method characterized in that the speech codec is generated by subjecting a noise code vector obtained from a book to a pitch period in a second pitch synchronization obtained by analyzing the adaptive code vector. 4. A speech decoding method for generating a synthesized speech approximating an input speech in a frame unit based on a drive signal comprising a linear sum of an adaptive code vector and a periodic noise code vector, wherein the adaptive code vector is inputted. Based on the first pitch period obtained from the index obtained from the adaptive codebook comprising the past drive signal, the periodic noise code vector is generated from the code vector obtained from the noise codebook based on the input index. A speech decoding method, wherein the speech signal is generated by subjecting the adaptive code vector to pitch periodization at a second pitch period obtained by analysis. 5. A correlation value based on the autocorrelation of the adaptive code vector is determined, and when the correlation value in a pitch period shorter than the first pitch period is equal to or greater than a predetermined threshold, a pitch period for providing the correlation value is determined. The second pitch period, the first pitch period;
4. The speech coding according to claim 1, wherein the first pitch cycle is set to the second pitch cycle when the correlation value at a pitch cycle shorter than the pitch cycle is less than the threshold value. Method. 6. A correlation value based on the autocorrelation of the adaptive code vector is obtained, and when the correlation value in a pitch period shorter than the first pitch period is equal to or greater than a predetermined threshold, a pitch period giving the correlation value is determined. The second pitch period, the first pitch period;
5. The speech decoding according to claim 2, wherein when the correlation value at a pitch cycle shorter than the pitch cycle is less than the threshold value, the first pitch cycle is set to the second pitch cycle. Method. 7. A correlation value based on the autocorrelation of the adaptive code vector is obtained, and when the correlation value at a pitch period equal to an integer fraction of the first pitch period is equal to or greater than a predetermined threshold, the correlation value is calculated. The given pitch cycle is the second pitch cycle, and when the correlation value at a pitch cycle that is an integral fraction of the first pitch cycle is less than the threshold value, the first pitch cycle is set to the second pitch cycle. 2. A pitch cycle.
Or the speech encoding method according to 3. 8. A correlation value based on the autocorrelation of the adaptive code vector is obtained, and when the correlation value at a pitch period equal to an integer fraction of the first pitch period is equal to or greater than a predetermined threshold, the correlation value is calculated. The given pitch cycle is the second pitch cycle, and when the correlation value at a pitch cycle that is an integral fraction of the first pitch cycle is less than the threshold value, the first pitch cycle is set to the second pitch cycle. 3. A pitch period.
Or the speech decoding method according to 4.