JPH0566800A - Speech coding and decoding method - Google Patents

Abstract

PURPOSE:To reduce the distortion of the front portion of the update frame. CONSTITUTION:Against the spectrum envelope information update frame, make the pitch information update frame, selected from an adaptive coding table, as a 1/2 subframe, synchronize them, make the length of the residual information update frame (noise coding table search frame) same as the length of the spectrum envelope information update frame and superimpose for every frame and shift them. The original voice is passed through an LPC inverse filter to form a residual waveform, using this, make plural virtual coding corresponding to the code selected from the adaptive coding table, quantize noise code using the above, obtain plural respective candidates, select the candidate, which has the minimum distortion with respect to the original voice, among all candidates so as to decide the noise code and after that, the pitch information is quantized.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a coding method for coding speech at low speed and high quality, and a decoding method for decoding the coded speech code.

[0002]

2. Description of the Related Art As a speech coding method for achieving high quality at low speed, M. R. Schroeder, B.A. S. At
Reference by al. (Reference 1), “Code-exited”
linear prediction (CELP):
high-quality speech at ve
"ry low bit rate", IEEE Pro
c. of International Conf. A
cost. , Speech & Signal Pr
cess. , Pp. 937-940 (1985) method and T.K. Moriya, M .; Reference by Honda (reference 2), "Transform Coding of
Speech using a weighted v
vector quantizer ", IEEE JS
elected Areas in Communic
ations, JSAC-6, pp. 425-431
Methods such as (1988) have been proposed and studied. These methods are called CELP methods and are 4.8 kb.
It is intended for a coding rate of about ps to 8 kbps and has been widely studied. The CELP method uses a model that constitutes a voice by using spectral envelope information of the voice, pitch information, and residual information. Therefore, the encoding side extracts / quantizes these three types of information, and the decoding side reproduces a voice waveform based on these three types of transmitted information.

Each of these pieces of information is extracted / quantized in units of frames or subframes obtained by dividing a frame into a plurality of frames, and this is repeated. FIG. 7 shows an example of the relationship between the speech waveform, the update frame of the spectrum envelope information, the update frame of the pitch information, and the update frame of the residual information in the conventional method. In this example, the frame update period of the spectrum envelope information is long and is twice the frame update period of other information. One of the characteristics of the conventional CELP method is that there are periodical times at which the boundaries of all information update frames coincide, and in FIG. 7, points a, b, and c correspond to this time.

[0004]

However, when the boundaries of the update frames coincide with each other, the distortion concentrates on the boundaries, and there is a high possibility that the deterioration particularly at the beginning of the next frame will increase. .. This is because the minimum distortion reference range in the quantization value determination is closed to the frame or subframe. The quantized value of the value near the center of the frame is determined based on the minimum distortion criterion of the frame from the relationship with the values in the front and the rear. However, the quantized value of the last sample of the frame is determined only from the relationship with the preceding value, and is determined irrespective of the value of the leading portion of the next frame, which is the subsequent value. Therefore, when viewed from the value of the leading portion of the next frame, a value that results in increased distortion may be selected as the quantized value, which is a problem of the conventional method.

[0005]

In order to solve the above-mentioned problems of the conventional CELP system in the present invention, for example, as shown in FIG. 1 corresponding to FIG. Subframes) should be configured so that they are adjacent to each other. By quantizing the entire same subframe shown in FIG. 1 at the same time, it is possible to realize the quantization taking into consideration the relationship before and after the signal in any part. For example, wrapped transform processing (for example, reference (reference 3) so that overlapping subframe signals are smoothly connected.
H. S. Malvar, “Lapped Transf
orms for Effective Transf
orm / Subband Coding ”, IEEE
Trans. Accout. , Speech, Sign
al Processing, vol. ASSP-3
8, pp969-978, June 1990).

In the conventional CELP system, in order to prevent the quantization noise from accumulating each time a frame is overlapped, the voice of the current frame is encoded so as to remove the influence of the quantization noise of the past frame. .. Specifically, it is configured to quantize the signal obtained by subtracting the influence of the quantization noise of the previous frame (zero input response of the synthesis filter of the current frame) from the voice of the current frame. In order to realize this configuration, it is necessary that all the quantization of the previous frame speech is completed before the quantization processing of the current frame speech. Conventional C in which all processing is divided for each frame
In the ELP method, the zero input response of the synthesis filter of the current frame can be calculated because all the quantization of the speech of the previous frame is completed before the quantization processing of the speech of the current frame, and the influence of the quantization noise of the previous frame can be calculated. It is possible to cancel it. However, in the present invention, as shown in FIG. 1, for example, all the quantization of the previous frame speech is not completed before the quantization processing of the current frame speech. As it is, the influence of the quantization noise of the previous frame cannot be canceled in the next frame, so the following is done.

At the time when processing of a certain frame is started, the information that has not been quantized in the information of the previous frame is a residual. It is assumed that the unquantized residual is quantized under an appropriate condition and the coding process of the frame is advanced. However, in the process of quantizing the cycle of the adaptive codebook and the quantizing of the shape of the residual (index search of the noise codebook), a plurality of suboptimal candidates are left in addition to the optimum candidate, and finally, among these, Select the combination that minimizes the quantization noise. In the conventional CELP, the quantization of the cycle of the adaptive codebook and the quantization of the shape of the residual are determined once.

As described above, the main features of the present invention are that the wrapped transform process is used, and a plurality of suboptimal candidates are left in the quantization process of the adaptive codebook cycle and the residual form. The final point is to select a combination that minimizes the quantization noise.

[0009]

FIG. 2 shows an embodiment of the decoding method according to the present invention. An envelope code indicating spectrum envelope information, a pitch code indicating pitch information, and a residual code indicating residual information are input as a reception code or a read output code of a file, and the residual code of the current one frame before is the first code. The noise waveform which is given to the random codebook 11 as an index of 1 and read from this is subjected to the inverse modified wrapped transform conversion in the converter 12, and the converted output having the two frame length is the latter half of the waveform in the latter half cutout circuit 13. Is cut out and given to the gain applying circuit 14, and its output is supplied to the adding circuit 15 as a first basic residual.

The residual code of the current frame is given as a second index to the noise codebook 16 which is the same as the noise codebook 11, and the noise waveform read from this is subjected to inverse modified wrapped transform conversion by the converter 17. The converted output having the two-frame length is cut out by the first half cutout circuit 13 into the first half of the waveform and is given to the gain giving circuit 19, and the output thereof is supplied to the addition circuit 15 as the third basic residual.

The input pitch code is given to the adaptive codebook 21, and the adaptive codebook 21 is supplied at the cycle indicated by the pitch code.
Are repeatedly read, and the read waveform is given a gain by the gain applying circuit 22 to obtain the second basic residual as the adding circuit 1
5 is supplied. The output of the adder circuit 15 is stored in the adaptive codebook and is supplied to the linear prediction / synthesis filter 23 as a drive signal. The filter coefficient of the synthesis filter 23 is controlled by the input envelope code. Synthesis filter 2
A decoded voice signal is obtained from 3.

FIG. 3 shows the timing relationship of signals in the outline of the decoding process of this embodiment. Since the residual information is encoded such that the previous frame and the current frame overlap each other by half, the third and first basic residuals are temporally continuous,
In addition, the residual information is constructed while overlapping the preceding and following frames. Therefore, in the present invention, the distortion at the beginning of the frame is small, which is significantly different from the conventional CELP method.

FIG. 4 shows a processing outline of an embodiment of the encoding method according to the present invention. In the normal CELP method, the lag (corresponding to the pitch period) of the adaptive codebook is determined, and then the index of the random codebook is determined. However, in the method of the present invention, the order is reversed, and the index of the random codebook is determined, and then the lag of the adaptive codebook is determined.
According to the present invention, in order to prevent the occurrence of a mismatch when the index of the random codebook determined first is combined with the adaptive codebook (the second basic residual) determined later, before the index determination of the random codebook is performed. In the present invention, a virtual second basic residual (an estimated value sequence of the second basic residual) is introduced into.

Since it is difficult to analytically obtain the optimum virtual second basic residual, a plurality of candidates for the virtual second basic residual are created, and the minimum distortion of the decoded speech signal is used as the evaluation criterion. It was configured to hit. Where the virtual second
Let N be the number of candidates for the basic residual of. Further, M noise code candidates are left for each candidate of the virtual second basic residual. That is, this embodiment adopts a configuration in which N × M decoded speech candidates are created and the candidate with the minimum distortion is selected.

FIG. 5 shows an example of a coding device to which the coding method of the present invention is applied. The input voice from the input terminal 25 is supplied to the virtual second basic residual creating unit 26, the LPC inverse filter 27, and the optimum value searching unit 28. N virtual second basic residuals are created from the virtual second basic residual creating unit 26 and are supplied to the noise code determining speech synthesizers 29 _{1 to} 29 _N. A residual waveform obtained by removing the feature of the spectrum envelope from the input speech is obtained from the LPC inverse filter 27, and this residual waveform is supplied to all of the noise code determination speech synthesis units 29 _{1 to} 29 _N.

In the noise code determining speech synthesizer 29 ₁ , the first virtual second basic residual and the LPC inverse filter 27 are used.
The difference between the residual residual waveform and the residual residual waveform is obtained by the subtracter 31. From this, the noise residual, that is, the latter half corresponding to the first basic residual and the first half corresponding to the third basic residual is obtained. This subtractor 3
The output of 1 is modified wrapped transform transform by the transform unit 32, and the transformed output is one of the same noise codebooks 34 as the noise codebooks 11 and 16 in FIG. 2 by the weighted vector quantization / inverse quantization unit 33. The noise code is quantized with weight. In this case, M candidates are obtained, and M indexes are obtained for each frame with respect to the residual waveforms of two frames, each half of which is the first basic residual and the first half is the third basic residual. , Each of the indexes is inversely quantized into a noise waveform and supplied to the speech waveform synthesizing units 35 _{1 to} 35 _M , synthesized into M speech waveforms and optimized value searching unit 2
8 is supplied. Noise code determined for speech synthesis unit 29 _2-2
9 _N is similarly constructed.

The above processing will be described in more detail below. Virtual Second Basic Residual Generation Method In the normal CELP coding method, the adaptive codebook of a certain frame has a signal in which the residual waveform of the previous frame is cut out and repeated in appropriate intervals. .. In the method of the present invention, the residual waveform of the preceding frame of a certain frame is the weighted sum of the first, second, and third basic residuals multiplied by a gain. However, the quantization of the third basic residual among these basic residuals is not completed (determined). Therefore, the residual waveform of the previous frame and the second basic residual of that frame generated from this cannot be determined.

Therefore, this second basic residual estimation sequence is generated by the following method. First, the LPC inverse filter 2
At 7, the residual left after removing the feature from the original speech before the previous frame of that frame (called the first residual) is generated. The first residual does not include quantization noise. That is, when the synthesis filter (coefficients that have not been quantized) are driven by the first residual, a speech waveform that matches the original speech is obtained. Next, the sum of the quantized first and second basic residuals of the previous frame is calculated, and this signal is quantized to the quantized residuals (first, second and third) before the previous frame. A signal (called a second residual) is generated by connecting the basic residuals) so as to be continuous in time. The difference between the second residual and the first residual is the target waveform of the third basic residual (third basic residual before quantization). Further, a weighted linear sum of the first residual and the second residual is obtained, and this is called a third residual. next,
A part cut out from the third residual or a waveform generated by temporally interpolating the third residual and then cut out directly or repeatedly is defined as a virtual second basic residual.

Here, the process of generating the virtual second basic residual waveform from the third residual will be described in detail. The cross-correlation between the third residual and the residual of the current frame is calculated, and the time difference (pitch lag) that maximizes the cross-correlation is obtained. A section of the length of the pitch lag is cut out from the temporally new portion of the third residual. The waveform is extended while repeating the cut section to make it the same length as the frame. This is a virtual second
Is the basic residual of. In order to generate N virtual second basic residuals, N pieces are left from the one that maximizes the pitch lag that increases the cross-correlation, and the virtual second basic is generated by the above method according to each pitch lag. Generate the residuals. Third and First Simultaneous Basic Residual Simultaneous Quantization Method Details of the noise code determining speech synthesizer 29 ₁ in FIG. 5, that is, the third and first basic residual simultaneous quantization methods and optimum synthesized speech. The waveform search is shown in FIG. A waveform obtained by subtracting the virtual second basic residual generated as described above from the residual waveform of the frame is generated by the subtractor 31. This waveform is used as a first basic residual, and is connected by a synthesizing circuit 36 so as to be temporally continuous after the basic residual obtained in the same manner before the quantization of the previous frame, that is, the third basic residual. This connection waveform is converted by the conversion unit 32 into a modulated wrapped transform (Modulated Lapped Tra).
nsform: MLT, for example, refer to Reference 3 for details)
It is converted by an algorithm to obtain a waveform in the frequency domain with one-half the number of data. This waveform is subjected to weighted vector quantization (Weighted Vector) by the quantizer 33a.
Quantization: WVQ, the details of which are described in Document 2) processing to quantize into a random code in one of the random codebooks 34. The waveform in the frequency domain is represented by one of the noise waveforms in the noise codebook 34 for each frame. At this time, a feature of the present invention is that a plurality (M) of candidates are left. Further, the following processing is performed for each of the M candidates that remain.

(A) The dequantizer 33b uses the random codebook 34
And (b) the inverse quantized output is inverse-transformed wrapped transform transformed by the transforming unit 37.
(C) The converted output is synthesized by the speech waveform synthesizer 55 into a speech waveform. By the above processing, decoded speech for M candidates is generated. As a result of the above-described processing in which the candidate having the smallest distance from the original speech waveform is selected , a total of N × M decoded speech candidates are generated. The search unit 28 calculates the distance between each candidate decoded speech and the original speech waveform, and determines the decoded speech with the smallest distance.

Finally, the second basic residual is redetermined. For the re-determination, the quantized third and first quantized decoded speech that minimizes the distance in the second basic residual determination unit 38 in FIG.
The second basic residual (the signal of FIG. 4) is determined using the basic residual of the same (the signal of FIG. 4). The encoding process of the present invention is summarized above with reference to FIG.

(1) A plurality of (N) virtual second basic residuals are generated from a residual waveform obtained by LPC inverse filtering the original speech waveform. (2) For each virtual second basic residual, obtain a set of a plurality (M) of quantized third and first basic residuals. (3) Using the distance between the decoded speech and the original speech waveform as a criterion for selection, the optimum candidate is determined from N × M decoded speech candidates.

(4) Determine the second basic residual using the quantized third and first basic residuals that give the determined optimal candidate. In FIG. 2, since the inverse MLT processing is performed at high speed, the noise codebooks 11 and 16 and the inverse MLT circuits 12 and 17 are each set to one, and the first half of the residuals of the two frames converted into each time domain is set to the current frame. Then, the latter half may be used in the previous frame. In the above description, the number of overlapping subframes is 1, but a plurality of subframes may be overlapped.

[0024]

As described above, according to the present invention,
Since the residual information, that is, the noise waveform is quantized while overlapping some frames, the increase in the quantization noise at the frame boundary is suppressed, the average quantization noise is reduced, and the quantization is reduced. Just a bit.

[Brief description of drawings]

FIG. 1 is a time chart showing an outline of an example of frame updating according to the present invention.

FIG. 2 is a block diagram showing an example of a decoder to which the decoding method of the present invention is applied.

FIG. 3 is a time chart showing a processing flow of the decoder shown in FIG.

FIG. 4 is a time chart showing an example of an encoding method according to the present invention.

FIG. 5 is a block diagram showing an example of an encoder to which the encoding method of the present invention is applied.

FIG. 6 is a block diagram showing a specific example of a noise code determination speech synthesis unit in FIG.

FIG. 7 is a time chart showing frame updating in a conventional speech encoding method.

Claims (4)

[Claims] 1. A speech signal is divided into frames, a characteristic parameter of the frame is extracted and encoded, a component represented by the encoded characteristic parameter is removed from the speech signal, and a residual is calculated, The residual is the first
In a speech coding method that expresses the sum of the second and third basic residuals and quantizes each of these basic residuals, before the quantization of the first basic residual of each frame, A virtual second basic residual is generated, a virtual residual is generated by subtracting the virtual second basic residual from the residual of the frame, and the virtual residual and information before the previous frame are generated. Originally, the quantized values of the first basic residual of the frame and the third basic residual of the previous frame are simultaneously determined, and then the final quantized value of the second basic residual is again determined. Characteristic speech coding method. 2. In the generation of the virtual second basic residual, a residual left after the feature is removed from the original speech before the previous frame of the frame (referred to as a first residual). The sum of the quantized first and second basic residuals of its previous frame, and the quantized residual of the previous previous frame (first, second and third basic residuals) to this signal. Sum of differences)
To generate a signal (referred to as a second residual) connected in time, and construct a signal (referred to as a third residual) by linear processing from the first residual and the second residual. 2. The speech coding method according to claim 1, wherein the virtual second basic signal is generated by using a waveform generated by directly or repeatedly extracting a part cut out from the third residual. .. 3. The first basic residual of the frame and the third basic residual of the frame preceding the frame are generated by using a wrapped transform process and are quantized. The speech encoding method according to claim 1 or 2. 4. An envelope code, a pitch code and a noise code are input, a noise code book is read by the noise code, the read noise waveform is extended to two frames, and the noise waveform extended in the current frame is read. The first half is taken out, the latter half of the noise waveform stretched in the previous frame is taken out, the waveform of the adaptive codebook is repeatedly read with the above pitch code, the output of the adaptive codebook and the first half of the stretched noise waveform, and the above stretched A speech decoding method in which the latter half of the noise waveform is added and the addition output is supplied as a drive signal to a synthesis filter whose filter coefficient is controlled by the envelope code to obtain a decoded speech signal.