JP3099836B2 - Excitation period encoding method for speech - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applied to code-driven linear predictive coding and vector-sum driven linear predictive coding using a noise codebook, and is a highly efficient speech for digitally encoding a speech signal sequence with a small amount of information. The present invention relates to an encoding method and a decoding method thereof.

[0002]

2. Description of the Related Art In order to efficiently use radio waves in a digital mobile radio communication system and to efficiently use a storage medium in a voice storage service, a high-efficiency voice coding method is used. Currently, as a method of encoding speech efficiently, the original speech is divided into sections called frames, which are about 5 to 50 ms, at regular intervals, and the speech of one frame corresponds to the envelope shape of the frequency spectrum and the envelope shape. It has been proposed to separate the information into two pieces of information, that is, a drive excitation signal for driving a linear filter to be driven, and encode each of them. In this case, as a method of encoding the drive excitation signal, the drive excitation signal is separated into a periodic component considered to correspond to the fundamental frequency (pitch cycle) of the voice and another component (in other words, an aperiodic component). Encoding methods are known. Code-Excited Linear Prediction Coding (CE)
LP) and Vector Sum Driven Linear Prediction Coding (Vector Sum
Excited Linear Prodiction Coding (VSELP) method.
About each technology, MRSchroeder and BSAt
al: "Code-Excited Linear Prediction (CELP): High-q
uality Speech at Very Low Bit Rates ", Proc.ICASSP '
85, 25.1.1, pp. 937-940, 1985, and IAGerson and
MAJasiuk: "Vector Sum Excited Linear Prediction
(VSELP) Speech Coding at 8 kbps ", Proc. ICASSP'90,
S9.3, pp. 461-464, 1990.

In these encoding methods, as shown in FIG. 3, a parameter representing an envelope shape of a frequency spectrum is calculated in an audio analysis unit 12 for an original audio input to an input terminal 11. Usually, a linear prediction method is used for this analysis. The linear prediction parameter is encoded by a linear prediction parameter encoding unit 13, the encoded output is branched, and decoded by a linear prediction parameter decoding unit 14, and the decoded linear prediction parameter is converted to a linear prediction synthesis filter. It is set as 15 filter coefficients.

In the adaptive codebook 16, the immediately preceding past excitation vector is cut out at a length corresponding to a certain period (pitch period), and the cut-out vector is repeated until the length of the frame becomes equal to the frame length. Is output. Noise codebook 1
7, 18 output time-series code vector candidates corresponding to the non-periodic components of the voice. As shown in FIG. 4, the noise codebooks 17 and 18 are usually based on white Gaussian noise, and various code vectors having a length of one frame are stored in advance independently of the input speech.

The time series vector candidates from the adaptive codebook 16 and the noise codebooks 17 and 18 are weighted by an adder 19 in multipliers 21 ₁ , 21 ₂ and 21 ₃ , respectively.
_1, g _2, g ₃ are multiplied, multiply outputs adding section 2
It is added by two. This added output is supplied to the linear prediction synthesis filter 15 as a drive excitation vector, and the synthesis filter 15 outputs a synthesized (reproduced) voice. The distortion of the synthesized speech with respect to the original speech from the input terminal 11 is calculated by the distance calculation unit 23, and the codebook search unit 2 is operated according to the calculation result.
4, a candidate having a different cutout length in the adaptive codebook 16 is selected, another code vector is selected from the noise codebooks 17 and 18, and the weights g ₁ , g ₂ , and g of the weighted addition unit 19 are further selected. ₃ is changed so that the distortion calculated by the distance calculation unit 23 is minimized. A periodic code indicating the cut-out length of the adaptive codebook 16 when the distortion is minimized, a noise code indicating each code vector of the noise codebooks 17 and 18, and a weight code indicating weights g ₁ , g ₂ and g _3. , Linear prediction parameter codes are output as encoded outputs and transmitted or stored.

For decoding, as shown in FIG. 5, the input linear prediction parameter code is
, And its prediction parameter is set to the linear prediction synthesis filter 27 as a filter coefficient. A time-series code vector obtained by cutting out the previous driving excitation vector from the adaptive codebook 28 in the cycle using the immediately preceding past driving excitation vector obtained up to that time and the input periodic code, and repeating this for the number of frames. Is output, and the code vector indicated by the input noise code is read from the noise codebooks 29 and 31 as time-series vectors, respectively. These time-series vectors are weighted in accordance with the weighting code input by the weighted addition unit 32, and then added, and the added output is supplied to the synthesis filter 27 as a drive excitation vector, and reproduced from the synthesis filter 27. Voice is obtained.

The random codebooks 29 and 31 are the same as the random codebooks 17 and 18 used for encoding. Only one or more noise codebooks may be used. In code-driven linear predictive coding, all code vectors to be candidates are directly stored in a random codebook. That is, if the number of code vectors to be candidates is N, the number of code vectors stored in the noise codebook is also N.

[0008] In the vector sum excited linear predictive coding, the noise codebook, as shown in FIG. 6, (referred to as a basic vector) all code vectors stored are read out simultaneously, multiplying unit 33 ₁ ~ 33 _M Are multiplied by +1 or −1 by the noise codebook decoder 34, and the multiplied outputs are added and output as an output code vector. Therefore, the number of output code vectors becomes 2 ^M by a combination of +1 and −1 by which each basic vector is multiplied, and one of the 2 ^M output code vectors is selected so as to minimize distortion.

However, in these conventional methods, since the periodicity of the driving sound source signal is limited to only the components of the previous frame, the expressiveness of the periodicity is weak, and the reproduced sound is rough and lacks smoothness. Had. From such a point, in order to enhance the expressive power of the periodicity of speech, a part or all of the code vector output from the noise codebook which did not have the periodicity in the past, or the component of the output code vector is It has been proposed that some or some of the noise codebooks have the same periodicity as that of the output time-series code vector of the adaptive codebook.

That is, as shown in FIG.
From the basic period search (adaptive code 1
6) is cut out for the length 36 of the basic period L obtained in (6). As shown in a, the cutout portion 36 is repeatedly arranged until the frame length is reached, and a periodic code vector is created to be an output code vector. This is performed for all the code vectors in the noise codebook 17, and among those, the one that minimizes the distance between the reproduced sound passed through the synthesis filter and the original sound is defined as the optimum code vector. Subsequent determination of the weight of each driving sound source component is performed in the same manner as in the related art. The decoding side also periodicizes the code vector of the noise codebook with the pitch period obtained so far.

[0011]

The quality of the reproduced speech can be improved by making the code vector of the random codebook periodic as described above. On the other hand, in the related art, the optimum period (pitch period) is determined using only the adaptive codebook as shown in FIG. 8, and then the index of the noise codebook, that is, the code vector is determined. It has been found that the period cannot be determined and, for example, a period that is twice the correct pitch period is often determined as the optimum period. For this reason, correct encoding may not be performed in some cases, and the encoding distortion cannot be reduced accordingly.

[0012]

According to the present invention, a code vector of a random codebook is adapted to a pitch period to perform a periodizing process, and iterative processing is performed in consideration of both vectors of an adaptive codebook and a random codebook. Determine the period, that is, the pitch period. By using a periodicized code vector of the noise codebook in this manner, the dependence between the code vector from the adaptive codebook and the code vector from the noise codebook becomes strong, and distortion in the frame is minimized. An optimal repetition period is determined. For this reason, the coding distortion can be further reduced as compared with the conventional method in which the pitch period of the adaptive codebook is obtained and is used as it is as the repetition period of the noise codebook.

[0013]

FIG. 1 shows a first embodiment of the present invention. In this embodiment, a loop for setting the repetition period and evaluating the distortion to determine the period includes a loop for searching for the index of the code vector of the noise component. That is,
The cycle is set within a predetermined pitch cycle range (existing range of a normal pitch cycle), and an excitation (drive) signal output from the adaptive codebook is created in the same manner as in the related art. Assuming this period and this excitation signal (excitation vector), each code vector of the noise codebook is subjected to a periodic process as shown in FIG. 7, and the periodic noise code vector is added to the excitation vector. The synthesis filter is driven, and an index search for a noise code vector that minimizes distortion is performed. This process finds the best index for the first set period,
That is, the index is provisionally determined. Thereafter, the same is repeated by changing the cycle to be set sequentially. Finally, a combination of the index and the cycle at which the distortion is minimized is obtained from the provisionally determined indexes.

As described above, since the adaptive codebook and the random codebook are used and the code vector of the random codebook is periodic, the dependence on the code vector of the adaptive codebook becomes strong.
Another cycle such as the double pitch does not become the optimum cycle. FIG. 2 shows a second embodiment of the present invention. In this embodiment, instead of searching for the index of the random codebook for all the determined repetition periods, preliminary selection of the period is performed, and index search is performed only for the preselected period. Preliminary selection is also used for the index search of the random codebook. In the first embodiment, optimal values are obtained for all combinations of the period and the index of the random codebook. However, since the search loop is doubled,
Depending on the conditions, the processing amount becomes very large. Therefore, in the second embodiment, the search is performed by narrowing down the period and the index to a small number of candidates.

As a preselection method of the period, the distortion is evaluated by using only the excitation signal from the adaptive codebook as in the prior art.
In addition to the minimum distortion value, a plurality of predetermined cycles are used in ascending order of distortion including the minimum value. Alternatively, a plurality of delay amounts for which the correlation function of the signal becomes large may be used as the period candidates. That is, conventionally, a delay amount for which the autocorrelation becomes large in order to obtain the pitch period is obtained. Since the distance calculation is not performed when the pitch period is obtained by the autocorrelation, the amount of calculation is significantly smaller than when the pitch period is obtained by the adaptive codebook search.

As a method of preliminary selection of the code vector (index) of the noise codebook, one cycle with minimum distortion is determined by the output of only the adaptive codebook, and each code vector of the noise codebook is determined at that cycle. Periodic, using the periodic code vector to find an index that minimizes distortion,
A plurality of indexes having a small distortion including this are set as candidates. With respect to the plurality of candidate indices of the noise codebook, the one that minimizes distortion with respect to another preselected period is obtained. Alternatively, one cycle is determined by the output of the adaptive codebook only, and the correlation between the error component excluding the component corresponding to the code vector of the applicable codebook and each noise code vector of the noise codebook is determined. Indices of some random code vectors may be used as preliminary selection candidates.

In the above description, the periodicization of the code vector of the random codebook is not performed for all code vectors (indexes), but may be performed only for a predetermined index. Further, the periodicization of the noise code vector may be performed not only at the period of the periodicization with respect to the code vector of the adaptive codebook, but also at a period twice or half the period. Further, the present invention can be applied not only to CELP but also to VSEL.
Also applies to P.

[0018]

As described above, according to the present invention, the repetition period used in both the adaptive codebook and the noise codebook can be optimally determined including the index of the noise codebook. Waveform distortion due to the above can be reduced. Also, by using the preliminary selection together, it becomes possible to obtain an almost optimal cycle with a practical range of processing amount.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a period determination processing method which is a main part of a first embodiment of the present invention.

FIG. 2 is a flowchart showing a period determination processing method using preliminary selection, which is a main part of a second embodiment of the present invention.

FIG. 3 is a block diagram showing a general configuration of a linear prediction encoding device.

FIG. 4 is a diagram showing a noise codebook in CELP.

FIG. 5 is a block diagram showing a general configuration of a linear prediction code decoding device.

FIG. 6 is a diagram showing a random codebook in VSELP.

FIG. 7 is a diagram showing periodicization of a code vector.

FIG. 8 is a flowchart showing a cycle determination processing method before improvement.

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kazunori Mano 1-6, Uchisaiwaicho, Chiyoda-ku, Tokyo Nippon Telegraph and Telephone Corporation (56) References JP-A-63-37724 (JP, A) JP-A Sho 64-40899 (JP, A) JP-A-64-54497 (JP, A) JP-A-2-84699 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/00 -19/14 H03M 3/00-11/00 B04B 14/00-14/08

Claims (1)

(57) [Claims] An audio signal is generated by driving a synthesis filter on a frame-by-frame basis using a time series vector obtained by repeating a past excitation vector from an adaptive codebook at a pitch cycle and a time series vector from a noise codebook. In a speech excitation cycle encoding method for encoding an input speech by using reproduction, a noise code vector of the noise code book is repeated every cycle corresponding to a repetition cycle of a time series vector from the adaptive code book. A time series vector from the noise codebook is obtained, and the synthesis filter is driven by a vector of the sum of the time series vector from the adaptive codebook and the time series vector from the noise codebook, and the speech signal is generated. Characterized in that the pitch period is determined so that distortion of the reproduced audio signal with respect to the input audio is reduced. .