JPH06195098A - Speech encoding method - Google Patents

Abstract

PURPOSE:To provide the speech encoding method which can obtain a synthesized speech of high quality even at a <=4Kbps low bit rate. CONSTITUTION:An input speech is analyzed by an acoustic classification part 31 and a retrieval code book selection part 34 selects a code book for retrieval processing among a pulse information code book 33, a pulse sound source code book constituted by a pulse generation part 34 and a pulse sound source retrieval part 32, a noise information code book, and a noise sound source code book constituted by a noise sound source retrieval part 37 according to the analytic result; and an in-use sound source selection part 35 selects a sound source to be used according to the retrieval result and the input speech is encoded corresponding to its acoustic features. As for pulse sound sources, a retrieval arithmetic quantity is reduced and all possible combinations of pulse sound sources are retrieved to select an optimum sound source pulse. Consequently, the reproducibility of a periodic component of a speech and adaption to acoustic features are improved and the speech of high quality can be obtained with a low throughput even at a low-bit rate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding method suitable for obtaining high-quality synthesized speech at a low bit rate, and particularly to a speech coding method applicable to a bit rate of 4 kbps or less with a relatively small processing amount. Pertain to.

[0002]

2. Description of the Related Art Recently, a speech coding method incorporating a "synthesis analysis" method for evaluating a weighted error between synthetic speech and original speech and determining a coding parameter so as to minimize the error has been proposed. We have succeeded in obtaining relatively good voice quality even at low bit rates. A typical example is a code driven linear predictive coding (CELP) method (for example, MR Schroeder and BS Atal: "Code-ex.
cited linear prediction (CELP) ", Proc. ICASSP 85 (1
985.3)) and achieves practical voice quality at 4.8 kbps. Also, many improved methods of the CELP method have been proposed, for example, vector sum driven linear predictive coding (VSELP) method (for example, IA Gersonand MA).
Jasiuk: "Vector sum excited linear prediction (VSE
LP) speechcoding at 8kbps ", Proc. ICASSP 90 (1990.
4)) is superior in processing amount, memory capacity, and bit error resistance.

On the other hand, with the full-scale digitization of mobile radio communications, there is a demand for the development of a voice coding system having a lower bit rate (4 kbps or less) from the viewpoint of effective use of frequencies. If CELP or VSELP is simply made to have a low bit rate, quality deterioration becomes large and there is a limit. This is because the long-term prediction accuracy by the adaptive codebook search is reduced and the reproducibility of the periodic component is reduced, resulting in a stronger sense of noise in the decoded speech. Therefore, in addition to the conventional statistical sound source (noise source), a method of introducing a pulse sound source to improve the reproducibility of periodicity has been proposed.

As such a method, a single pulse whose phase and amplitude are controlled is used for voiced sound, and CELP is used for unvoiced sound.
"SPE-CELP" method (W. Granzow and
BSAtal: "High-quality digital speech at 4 kb /
s ", Proc. GLOBECOM 90 (1990.12)), or" pulse / noise selection type CEL "that switches between periodic pulse and noise
P "method (Yoshida et al. 2:" Application of pulse source search to low bit rate CELP coding ", IEICE Technical Report SP91-
68 (1991.10) or Tanaka, Itakura: "C
Improving Speech Quality by Introducing a Pulse Sound Source in ELP Speech Coding System ", IEICE Technical Report EA92-24 (199)
2.5)) etc.

[0005]

The speech coding method using the above pulse sound source can improve the reproducibility of the periodic component even if the bit rate is reduced as compared with the conventional method, but has the following problems. .

Since the "SPE-CELP" method uses only one pulse per pitch period, its position and amplitude have a great influence on voice quality. The method of determining the pulse position is quite complicated, and there is a problem in robustness to the input audio signal. In addition, there is also a report that the coded voice may become a buzzer.

On the other hand, the "pulse / noise selection type CELP" system evaluates an error when a pulse sound source and a noise sound source are individually used, selects a sound source with a smaller error, and determines whether a voiced voice is unvoiced. Select the sound source to use. Since these methods also use long-term prediction (adaptive codebook search), pulse sources have a strong meaning of complementing long-term prediction vectors. However, in the above-mentioned document, since the pulse interval is limited to the long-term prediction lag or the pitch period, there is a problem that sufficient voice quality is not obtained.

Further, in both the "SPE-CELP" system and the "pulse / noise selection type CELP" system, switching between the pulse sound source and the noise sound source is performed, and therefore, the tone color change caused by the sound source switching in the coded voice. There is also the problem of (discontinuity).

Further, in the method using only the pulse component as the sound source, it is difficult to approximate the residual waveform by the pulse sound source in a portion where the residual waveform such as a fricative noise becomes noisy, and the deterioration of the decoded speech is conspicuous. It is problematic to use only pulsed sound sources.

A first object of the present invention is to provide an encoding method in which the deterioration of voice quality is small even if the bit rate is reduced. A second object of the present invention is to realize the first object with a relatively low throughput.

[0011]

In order to achieve the above object, in the present invention, the speech input to the speech coder is first divided into frames and subframes. In the short-term prediction analysis unit, the spectrum parameter (short-term prediction coefficient) is extracted and quantized for each frame. The input speech is then perceptually weighted in preparation for evaluating perceptual weighting errors. Further, the zero signal is input to the weighting synthesis filter, the zero input response is obtained, and the zero input response is subtracted from the weighted input signal. This is to remove past effects that depend on the internal state of the synthesis filter. Furthermore, the impulse response of the weighting synthesis filter is also calculated.

Next, the long-term prediction analysis unit obtains the optimum long-term prediction lag and gain from the adaptive codebook in subframe units. A signal obtained by subtracting the weighted long-term prediction vector multiplied by the gain from the signal obtained by subtracting the zero-input response from the weighted input signal is input to the search codebook selection unit.

The sound classification unit analyzes the input speech in units of frames or subframes, obtains acoustic parameters representing acoustic characteristics, and outputs the analysis result to the search codebook selection unit and the used sound source selection unit. (A) The search codebook selection unit selects a codebook to be searched from a plurality of codebooks based on the input information from the acoustic classification unit, and inputs the above-mentioned search target signal to each search codebook. The codebook has multiple sound sources with different characteristics such as pulse sound source and noise sound source.
An appropriate codebook is selected for the search process based on the acoustic features of the input speech. (B) In the search for the pulse sound source, first, the information on the pulse interval and the head pulse position is read from the pulse information codebook, and the pulse train is generated by the pulse generator. At this time, the information on the pulse interval is not limited by the search results of long-term prediction,
All preset values are used for pulse train generation. This pulse train is regarded as a sound source vector and weighted by convoluting the impulse response of the weighting synthesis filter. The weighting error is sequentially evaluated for these weighting vectors, and the pulse source vector and the gain that minimize the error are determined. (C) In the noise source search, noise information is read from the noise information codebook, a source vector is created, and the weighting error of the weighting vector obtained by weighting the source vector is evaluated to minimize the error. Determine the noise source vector and gain. (D) The used sound source selection unit selects a used sound source codebook from the search result of each search codebook selected by the search codebook selection unit, the analysis result of the acoustic classification unit, the search result of the long-term predictor, and the like. The search result of the codebook is output as a sound source vector, and the used sound source index indicating the codebook to be used is output. (E) The gain quantizer simultaneously optimizes and quantizes the gains of the long-term prediction vector and the excitation vector.

The spectrum parameter, the quantized code of the gain, the long-term prediction lag, the used excitation index, and the excitation vector index obtained as described above are transmitted to the decoder as transmission parameters.

In the decoder, the driving sound source is calculated from the above transmission parameters and is input to the synthesis filter using the short-term prediction coefficient as a filter coefficient, whereby decoded speech is obtained.

[0016]

In the acoustic classification section (A), the input voice is analyzed to obtain the parameter representing the acoustic feature, and the acoustic classification is performed based on the parameter. Then, the acoustic parameters of the input voice and the classification result are output to the search codebook selection unit and the used sound source selection unit.

The search code book selecting section in the above (B) is
According to the analysis result of the sound classification unit in (A) above, the codebook to be searched is selected and limited. This can reduce the amount of calculation required for codebook search,
Moreover, the quality of the synthesized voice is maintained by selecting an appropriate codebook for the input voice.

As a sound source codebook, a plurality of codebooks having different characteristics such as the pulse sound source of (C) and the noise sound source of (D) are prepared, so that the acoustic characteristics of the input voice such as a stationary portion and a non-stationary portion. It is possible to cope with the difference of the above, prevent the deterioration of the approximation accuracy of the driving sound source, improve the coding efficiency, and improve the sound quality of the synthesized voice.

In the pulse sound source of the above (C), the start pulse position of the pulse train to be searched and the range of the pulse interval are determined in advance, and the entire range is searched regardless of the result of the long-term prediction or the pitch prediction. It improves the approximation accuracy of the sound source, compensates for the deterioration of the long-term prediction gain and the reproducibility of the periodicity due to the low bit rate, and improves the sound quality of synthesized speech. In the pulse sound source search process, the impulse response of the synthesis filter is cut off and the minimum pulse interval is set to eliminate the influence between adjacent pulses, thereby reducing the amount of calculation. Further, the pulse information codebook for generating the pulse train does not have the information of the waveform vector itself but only the two information of the head pulse position and the pulse interval, so that the memory amount required for the codebook can be reduced. .

The used sound source selection unit (E) selects a used sound source from the selection result of the search codebook selection unit, the search result of each codebook, the analysis result of the sound classification unit, and the like. This makes it possible to select an optimum driving sound source and improve the approximation accuracy of the driving sound source.

[0021]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows a block diagram of a speech encoding unit according to the embodiment of the present invention, and FIG. 2 shows a block diagram of a speech decoding unit.

The present invention uses code driven linear prediction (CELP).
Since it is based on the voice coding system, the principle of the CELP system will be described first before the description of a specific embodiment. FIG. 3 is a principle diagram of driving sound source determination in the CELP coding unit. In the figure, the long-term prediction vector 110 that is the output of the adaptive codebook 108 as a component that represents the periodicity of the sound source, and the code vector 111 that is the output of the statistical codebook 109 as the components other than the periodicity (randomness and noise). Is added to each of the gains 112 and 113 and added as a driving sound source.

The codebook search for obtaining the optimum driving sound source is performed as follows. Generally, it suffices to obtain a driving sound source in which the synthetic speech obtained by inputting the driving sound source to the synthesis filter matches the original speech (input speech), but in practice, some error (quantization distortion) is involved. Therefore, it is sufficient to determine the driving sound source so as to minimize this error, but it is known that the human auditory characteristics do not always correspond to the error amount and the subjective quality of voice. Therefore, it is general to use an error weighted so that the correspondence with the auditory characteristics is improved. Hearing weighting is described in the following documents, for example. BS Atal
and JR Remde: "A new model of LPC excitation f
or producing natural-sounding speech at low bit ra
tes ", Proc. ICASSP 82 (1982.5).

To evaluate this perceptual weighting error,
The driving sound source 114 is input to the weighting synthesis filter 105 to obtain the weighting synthesis speech 116. The input voice 101 also obtains a weighted input voice 115 through the auditory weighting filter 104, and obtains a weighted error waveform 117 by subtracting the weighted input voice 115 from the weighted synthetic voice 116. Note that the filter coefficients of the perceptual weighting filter 104 and the weighting synthesis filter 105 are the LPC (linear prediction) of the input speech 101 in advance.
LPC parameter 10 obtained by inputting to the analysis unit 102
Determined by 3.

The weighted error waveform 117 is calculated as the sum of squares over the error evaluation section in the squared error calculation unit 118, and the weighted squared error 119 is obtained. Since the driving sound source is the weighted sum of the long-term predicted vector and the statistical code vector as described above, the determination of the driving sound source results in the determination of the code vector index that determines which code vector is selected from each codebook. That is, the long-term prediction lag 106 and the code vector index 107 are sequentially changed to calculate the weighted squared error 119, and the error minimizing section 120 may select the one with the smallest weighting error. Such a driving sound source determination method is called a "synthesis analysis" method.

When the optimum driving sound source is determined in this manner, the long-term prediction lag 106, the codebook index 107, the gains 112 and 113, and the LPC parameter 103 are multiplexed as transmission parameters in the multiplexing unit 121, and the transmission data 122 is obtained. To do. Also, the driving sound source 114 at this time
Is used to update the state of adaptive codebook 108.

If the above-mentioned "analysis by synthesis" method is faithfully executed, that is, if the long-term prediction lag and the index of the statistical code vector are simultaneously optimized while evaluating the weighting error, a huge amount of processing is required. . Therefore, a method such as sequential optimization is actually used.

On the other hand, in the processing in the decoding section, the received data 222 is first separated into various parameters by the demultiplexing section 221. The adaptive codebook 208 is searched based on the long-term prediction lag 206, and the long-term prediction vector 210 is output.
Also, the statistical codebook 209 is searched based on the codebook index 207, and the sound source vector 211 is output. The long-term prediction vector 210 and the sound source vector 211 are multiplied by respective gains 212 and 213, and the added signal is input to the synthesis filter 230 as the driving sound source 214. The filter coefficient of the synthesis filter is determined by the LPC parameter 203. Although the post filter 231 is not essential, it is often used to improve the subjective quality of synthesized speech, and its output becomes the output speech 232.

FIG. 1 shows a block diagram of a speech coder according to an embodiment of the present invention, and FIG. 2 shows a block diagram of a speech decoder. The outline of the operation of this embodiment will be described below.

In the voice encoding unit, the digital voice signal 11 A / D converted at a predetermined sampling frequency (usually 8 kHz) is input.

The short-term prediction / analysis unit (LPC analysis unit) 12 reads out the voice data 11 having the analysis frame length and outputs the short-term prediction coefficient 13. The frame length is, for example, 40 ms (3
20 samples).

The short-term prediction coefficient 13 is quantized in the short-term prediction coefficient quantizer 14. The quantized code is output as the transmission parameter as the short-term prediction coefficient quantization index 18. Further, the quantized value 17 of the short-term prediction coefficient is referred to in the processing of the next and subsequent stages.

Further, the input voice 11 is weighted by the perceptual weighting filter 19 to obtain the weighted voice 20. On the other hand, a signal (zero input) 22 having a value of 0 for the frame length is input to the weighting synthesis filter 21, and a zero input response 23 is obtained. Subtract this from the weighted input voice 20,
The weighted input speech 24 is obtained by removing the influence of the past internal state of the weighting synthesis filter. Further, the impulse response 29 of the weighting synthesis filter is also obtained.

Since the long-term predictive analysis is performed by searching the adaptive codebook for each subframe, it will be referred to as an adaptive codebook search hereinafter. Here, the subframe length is, for example, about 10 ms (80 samples). The adaptive codebook search unit 25 extracts the long-term prediction lag, which is a parameter indicating the periodicity of speech, and outputs the long-term prediction lag index 30 and the long-term prediction vector 58.

The sound classification unit 31 analyzes the input voice 11 in units of frames and subframes, and outputs the sound classification parameter 33 to the search codebook selection unit 34 and the used sound source selection unit 35.

The search codebook selection unit 33 selects a search target codebook from a plurality of codebooks according to the sound classification parameter 33 from the sound classification unit 31, the analysis result up to the previous frame, and the coding result. On this occasion,
A part of each codebook is also regarded as a codebook and is searched. By limiting the search target codebook in this way, the approximation accuracy of the driving sound source is maintained while reducing the amount of calculation required for code vector search.

When the search codebook selecting section 34 selects the pulse sound source, the pulse generating section 40 reads the information 39 on the pulse interval and the leading pulse position from the pulse information codebook 38, and generates the pulse train 41 based on the information.
In the pulse sound source search unit 36, the pulse train 41 is regarded as a sound source vector, and the impulse response 2 of the weighting synthesis filter is calculated.
Weighting is performed by convolution of 9. A weighted long-term prediction vector 28 obtained by further multiplying a gain from a signal 24 obtained by subtracting a quiescent response 23 from a weighted input signal 20.
The optimum pulse source vector 4 for the signal from which
Search for 6. The index 44 of the pulse information codebook 38 corresponding to the optimum pulse source vector 46 is output.

When the search codebook selecting unit 34 selects a noise source, the noise source searching unit 37 reads the noise information 43 from the noise information codebook 42, generates a noise vector, and regards this noise vector as a source vector. ,
The same filter coefficient as the weighting synthesis filter 21 is calculated from the quantized value 17 of the short-term prediction coefficient, and the weighting is performed by the coefficient. Then, the optimum noise source vector 47 is searched for the signal 24 obtained by subtracting the zero input response 23 from the weighted input signal 20 and further subtracting the weighted long-term prediction vector 28 multiplied by the gain. The index 45 of the noise information codebook 42 corresponding to the optimum noise source vector 47 is output.

In the used sound source selection unit 35, the sound classification unit 31
From the sound classification parameter 33, the selection result 32 of the search codebook selection unit 34, and the search results 44, 45, 46, and 47 of the search target codebook, the codebook index 32 of the sound source, the sound source code vector 50, and the sound source code. The vector index 49 is output.

The gain optimization / quantization unit 51 calculates and quantizes the optimum values of the gains of the long-term prediction vector 58 and the excitation vector 50. The quantized code 52 at that time is output.

The short-term prediction coefficient and gain quantization code 18, 52, the long-term prediction lag index 30, the used sound source index 32, and the sound source information codebook index 49 obtained as described above are used as transmission parameters in the speech decoding unit. Transmitted to.

In the speech decoding unit, the long-term prediction lag index 6
3 is used to read the long-term predicted vector 69 from the adaptive codebook 68, and the pulse generator 73 uses the sound source codebook index 64 to read information 71 on the pulse interval and the head pulse position from the pulse information codebook 70. Generate pulse source vector 74. The noise information codebook 75 uses the source codebook index 64 to generate a noise source vector 76. Then, the sound source selection unit 70 switches the sound source, and the sound source vector 77
Is output. Also, using the gain codebook index 66, each gain 79, 80 is reproduced from the gain codebook 78. Gains 79, 8 for each code vector 69, 77
The driving sound source vector 84 is generated by multiplying by 0 and adding.

By inputting the driving sound source 84 to the synthesis filter 85, a synthetic voice 86 is obtained. The filter coefficient of the synthesis filter 85 is the short-term prediction parameter 8 read from the short-term prediction parameter quantization codebook 81 based on the quantization index 67 of the short-term prediction parameter.
2 is used. Finally, for the purpose of improving subjective sound quality, the synthetic speech 86 is input to the adaptive post filter 87, and the final decoded speech 88 is obtained.

The decoded voice (digital signal) is DA converted, converted into analog voice and output.

The outline of the present embodiment has been described above. Next, the detailed functions of the main parts will be described.

The short-term prediction analysis unit (LPC analysis unit) 12 is
The short-term prediction coefficient 13 representing the spectrum envelope of the voice is extracted from the voice data 11 for each frame. Short-term prediction coefficient 1
3 is most commonly a linear prediction coefficient, but is an equivalent parameter derived from it, the partial autocorrelation coefficient (PA
It is easily converted into RCOR coefficient, reflection coefficient) and line spectrum pair (LSP parameter).

As a method of deriving the linear prediction coefficient, Dur
Bin-Levinson Iterative Method (Saito, Nakata,
"Basics of voice information processing", introduced by Ohmsha, Ltd. in 1981) is common, and the derivation method of the reflection coefficient is
In addition to the above, the FLAT algorithm (defined by the Radio System Development Center, "Digital Car Telephone System Standard RCR STD-27" (hereinafter "RCR Standard")
Abbreviated) and the LeRoux method (Saito,
Nakata, written in the above-mentioned book) is proposed. Also,
The conversion method from the linear prediction coefficient to the LSP parameter is also described in the above-mentioned book by Saito and Nakata.

In the present embodiment, the linear prediction coefficient 13 is converted into an LSP parameter and then vector-quantized by the quantizing unit 14 to be converted into a quantized value 17 (the code vector 16 is sequentially read from the LSP codebook 15). The quantized value is the one with the smallest error). LSP
It is known that the parameter has a better quantization characteristic than that of directly quantizing the linear prediction coefficient (spectral distortion is small even if quantized with the same number of bits). The quantization method is
Depending on the number of bits allowed, scalar quantization, multistage vector quantization, vector scalar quantization, etc. may be used. The quantization index 18 is output as a transmission parameter.

Next, preprocessing for calculating the perceptual weighting error will be described. In order to calculate the weighting error, the perceptual weighting filter 19 first weights the input voice 11 to obtain a weighted voice 20. The weighting filter 19 is a short-term prediction coefficient (or equivalent parameter).
Of the quantized value 17 of, the concrete form of which is as follows.

[0050]

[Equation 1]

Here, α _i is a filter coefficient (linear prediction coefficient), Np is a filter order, for example, Np = 10, and λ is a weighting parameter, usually λ = 0.8.

Generally, the output of the synthesis filter is influenced by the past state, but here, in order to reduce the amount of calculation, the influence of the past synthesis filter is removed from the weighted speech 20 in advance. That is, data having a value of 0 corresponding to the frame length in the weighting synthesis filter 21 (zero input 22)
, The zero input response 23 is calculated, and the weighted speech 20
Weighted speech 24, which is subtracted from
To get The transfer function of the weighting synthesis filter 21 used here is as follows.

[0053]

[Equation 2]

This synthesis filter 21 is different from the decoding-side synthesis filter in that it includes the weighting parameter λ. Further, the impulse response 29 of the weighting synthesis filter 21 is also obtained at the same time. At this time, (Equation 2)
As α of, the quantized value 17 of the linear prediction parameter is used.

As explained at the beginning, the long-term prediction analysis is regarded as a search of the adaptive codebook, and the long-term prediction lag (index of the adaptive codebook) is selected by minimizing the auditory weighting error between the synthetic waveform and the original speech. . Here, the case where the adaptive codebook search and the pulse sound source search are sequentially performed will be described. That is, the optimum long-term prediction lag index 30 is determined on the assumption that a pulsed sound source is not used.

Next, the adaptive codebook search unit 25 will be described. The weighted synthesis of the code vector 27 read from the adaptive codebook 26 corresponding to the long-term prediction lag to be searched is realized by convolution with the impulse response 29 of the weighted synthesis filter. The synthesized output (weighted long-term predicted vector) 28 thus obtained does not depend on the past states of the synthesis filter, and is called a zero-state response. The long-term prediction vector 28 for each lag in the search range is calculated, the correlation with the weighted speech 24 from which the past influence is removed is calculated, and the (optimal) long-term prediction vector 58 that gives the maximum value of the correlation and the long-term prediction time at that time are calculated. The long-term prediction lag index 30 obtained by quantizing the prediction lag is output. For details of the long-term prediction analysis method and the method for reducing the amount of calculation, refer to the above-mentioned RCR standard.

Next, the combined use of the pulse sound source and the noise sound source and the generation of the pulse sound source vector will be described.

In the present invention, a plurality of codebooks are provided in place of the conventional CELP statistical sound source, and at least one pulse codebook and at least one noise codebook are provided in each of them, and the pulse codebook and the noise codebook are used together. The feature of the codebook is that all searches are performed independently of the results of long-term prediction analysis. The combined use of the pulse sound source and the noise sound source is performed because the residual waveform that the driving sound source is trying to approximate also has different properties depending on the acoustic properties of the input voice. For example, as shown in FIG. 5, the speech waveform can be roughly divided into a stationary portion and a non-stationary portion, but the residual waveform is obtained by removing a component having a short-term characteristic from the speech waveform and is a long-term periodic component. It can be considered that the characteristics of the residual waveform are different between the stationary part and the non-stationary part. It is difficult to approximate residual waveforms having different properties with only a pulse sound source or a noise sound source. Therefore, the sound source codebook to be searched is changed according to the acoustic characteristics of the input voice to improve the sound quality of the encoded voice.

In the present invention, the sound source to be used is selected by analyzing the acoustic characteristics of the input voice and classifying the sound, a sound codebook selecting unit 34 for selecting a sound source codebook to be searched, and a sound source to be used. This is performed by the used sound source selection unit 35 that selects a codebook. In the present embodiment, a plurality of search target codebooks are selected by the search codebook selection unit 34, and the used sound source selection unit 35 selects one codebook as the used sound source. By combining the selection methods of the respective sound source selection units 35, several sound source selection methods other than this embodiment can be considered. In selecting the search codebook, a method of selecting all or some of the codebooks or a method of including a subset of each codebook in the search codebook can be considered. It is also possible to limit the search codebook to one in a top-down manner depending on the output result of the sound classification unit, etc. In this case, the selection codebook selection unit also serves as the sound source selection unit, and the search codebook selection The result automatically becomes the used sound source index. In addition, in the method for selecting and evaluating the search sound source, a method of selecting based on the result of acoustic classification performed in units of frames or subframes, a method of selecting from the results of encoding up to the previous frame, and combining them A method of making a selection can be considered. When selecting the sound source to use,
A method of selecting a sound source to be used from a search result of a plurality of codebooks in a bottom-up manner, a method of selecting a sound source of a plurality of codebooks in consideration of a result of an acoustic classification unit, and the like are possible.

The pulse sound source is basically a part of the periodic pulse train (for the subframe length). However, as shown in FIG. 6, the head pulse position is set so that it can be taken from the first sample to the last sample of the subframe regardless of the pulse interval. This is because the sub-frame length becomes longer as the bit rate becomes lower, and the characteristics of the rising edge of the voice, which cannot be covered by the long-term prediction vector, are reproduced by the pulse sound source. Further, it is preferable that the pulse interval substantially covers the variation range of the pitch period of human utterance, similarly to the search range of the long-term prediction lag. In this embodiment, the minimum pulse interval is Lmi
n = 20 and the maximum pulse interval is Lmax = 146.

As shown in FIG. 7, the pulse information codebook 38 stores pulse intervals and head pulse positions. As can be seen from FIG. 6, when the pulse interval is L and the subframe length is N (N = 80 in this embodiment), L ≧
In the case of N, the number of pulses in the subframe is one. L
In the case of <N, the number is one or two or more depending on the head pulse position. Since the case of one line overlaps with the case of L ≧ N, the pulse interval and the leading pulse position are arranged in the pulse information codebook so that the pulse trains do not overlap. That is, when L <N, the leading pulse position is set to a range where two or more pulses are present in the subframe, and L
For ≧ N, L = N is representative and the head pulse position is from 0 to N−1. In this embodiment, N = 80, Lmin
= 20, so the number of pulse trains that do not overlap is 1
There are 910 types, but by setting the head pulse position every 2 samples, the number of types of pulse trains becomes 1010 types, which can be expressed by 10 bits. This is intended to reduce the number of transmitted bits, but as a result of experiments, the decoded speech 88 is less deteriorated and there is no problem in the performance of the speech coding unit.

The pulse generating section 40 generates a pulse as shown in FIG. 8 based on the pulse interval and head pulse position information 39 read from the pulse information code book 38. The amplitude of the pulse is 1, and the amplitude of the sample without the pulse is 0.

The above is the case where the pulse information codebook 38 and the pulse generator 40 generate the pulse excitation vector 41, but it is of course possible to store all the pulse excitation vectors in the codebook. However, in this case, the pulse generation process can be omitted, but the wordbook requires a storage capacity of N words, whereas the pulse information codebook 38 requires only two words of the pulse interval per vector and the head pulse position. become.

Next, the search for the pulse sound source will be described.

First, the optimum long-term prediction vector 58 output as a result of the adaptive codebook search is b _L (n), and its weighted signal (zero-state response of b _L (n)) 28 is b ′ _L (n). ),
The gain is β. Further, the weighted input voice 24 from which the influence of the past is removed is defined as p (n). Here, p '
Define (n).

[0066]

[Equation 3]

This represents the component obtained by subtracting the contribution of the long-term predicted vector from the ideal synthetic speech, which is the component to be covered by the pulse sound source.

If the generated pulse sound source is f _i (n) and the weighted synthesized speech is f ′ _i (n), the error E,

[0069]

[Equation 4]

It suffices to find f ′ _i (n) that minimizes Here, γ _i represents the gain, and i represents the index (index) of the pulse information codebook.

When (Equation 4) is partially differentiated by γ and set to 0, γ _i that minimizes the error E is

[0072]

[Equation 5]

And E at this time is

[0074]

[Equation 6]

It becomes 'Since the positive constant value irrespective of the _i (n), to maximize the second term f' where the first term on the right side of equation (6) is f _i (n), i.e. pulse excitation f _i ( Reduce to finding n).

The above-mentioned processing is basically performed by the conventional CELP.
It is the same as the statistical codebook search in, and is a large processing amount. In the present invention, by utilizing the characteristics of the pulse sound source and using the impulse response whose order is truncated, the amount of search processing is significantly reduced.

Generally, when synthesizing a voice by convolving the impulse response, the order cutoff of the impulse response causes an error. However, the impulse response of the weighting synthesis filter represented by (Equation 2) has a steeper attenuation than the impulse response without weighting as shown in FIG.
The impact of order cancellation is small. The censoring order is the 20th order (2.
If it is set to about 5 ms), in most cases the effect of censoring can be ignored. Therefore, in the present invention, the cutoff order is set to Lmin (20 samples) which is the minimum pulse interval of the pulse sound source.

Here, C _i and G _i are defined as follows.

[0079]

[Equation 7]

[0080] C _i is the cross-correlation p '(n) and f' _i (n),
The 'because it is the power of _i (n), would otherwise f' G _i is f _i
It is necessary to recalculate each time (n) changes (every time the index i is updated). On the other hand, p ′ (n) (0 ≦ n ≦ N−1, N
Is the number of subframe samples) and impulse response h (n)
Is constant in a certain subframe. Where the order is Lm
The impulse response truncated by in is h ′ (n) (0 ≦ n ≦ L
min) and a _j (0 ≦ j ≦ N−1) represented by the following equation
Is calculated in advance.

[0081]

[Equation 8]

As shown in FIG. 10, a _j represents the cross-correlation with the part of p '(n) corresponding to h' (n) when the position of h '(n) is shifted by one sample. I am doing it.

Since h '(n) is cut off by Lmin, no overlap occurs between pulses for any pulse sound source to be searched. Accordingly, the a _j To obtain the C _i, for example, as shown in FIG. 11, when the pulse position of the pulse excitation f _i (n) is that it was P1, P2, P3, which is previously calculated in equation (7) Of these, the sum of a _P1 , a _P2 , and a _P3 should be calculated. Therefore,
Since the calculation of the convolution of the impulse response that should be performed each time f ′ _i (n) changes is replaced with the sum of the partial cross-correlations that have been calculated once for each subframe, the processing amount can be significantly reduced. became.

A similar method can be applied to G _{i in} (Equation 7). That is, g _j defined in advance by the following equation
Is calculated.

[0085]

[Equation 9]

As shown in (Equation 9), 0 ≦ j ≦ N
In the case of −Lmin, the value of g _j is constant, so only g ₀ needs to be calculated. The calculation of G _i is similar to the case of C _i ,
This can be realized by _obtaining the sum of g _j corresponding to the pulse position of f _i (n).

When the optimum pulse sound source f _i (n) (maximizing C _i ² / G _i ) is obtained in this way, the impulse response h (n) without order censoring is used.
f _i strict weighting signal f _'i of (n) (n) previously calculated.

In the conventional method using the pulse codebook (the above-mentioned documents of Yoshida et al. And Tanaka et al.), The pulse interval is the long-term prediction lag or the pitch period obtained by pitch extraction. Therefore, the sound quality is deteriorated when the pulse sound source is used in the portion where the periodicity of the input sound is low. In the present invention, since all possible combinations of pulsed sound sources are searched, even in such a portion, the long-term predicted vector is complemented and good sound quality can be obtained.

In this embodiment, a binary weighted linear combination of a small number of basis vectors is used as a noise source, and a noise source search is performed using the binary weighted linear combination of a small number of basis vectors. It does this by searching for combinations. This method is used in the full rate speech coding method of the RCR standard, and has been greatly improved in terms of processing amount and required memory amount. Refer to the RCR standard for the actual processing procedure.

The final stage processing in the speech encoding unit is gain optimization and quantization. Gain optimization / quantization unit 51
, The strictly weighted long-term prediction vector b ′ _L (n) 28 and the source vector f ′ _i (n) 50 (determined by convolution of the impulse response without order truncation), and the past effects are removed. The weighted input voice p (n) 24 is input. Here, assuming that the gains are β and γ again, β and γ are determined so as to minimize the weighting error E of the following equation.

[0091]

[Equation 10]

Specifically, it is based on solving a simultaneous equation that can be set to 0 by partially differentiating (Equation 10) with β and γ.

As for the quantization of gain, there are methods such as direct scalar quantization of β and γ, scalar quantization after conversion into another variable, or vector quantization. In this embodiment, the latter method is used for scalar quantization.

When the quantized values of β and γ are β _q 53 and γ _q 54, respectively, the unweighted long-term prediction vector 50 and the excitation vector 50 are multiplied to produce a driving excitation 55. The driving sound source 55 is used to update the adaptive codebook 26.

Next, returning to FIG. 2, the speech decoding unit of this embodiment will be described.

The received data 61 is demultiplexed by the demultiplexing unit 62 into a short-term prediction parameter quantization index 67, a long-term prediction lag index 63, a used sound source index 65, a sound source information codebook index 64, and a gain quantization index 66. .

The first stage of the decoding process is the decoding of each parameter value. The short-term prediction parameter value 82 is decoded from the short-term prediction parameter quantization codebook 81 based on the short-term prediction parameter index 67. Similarly, the adaptive codebook 68 decodes the long-term prediction vector 69 based on the long-term prediction lag index 63. Gain Codebook 7
8, the quantization gain 7 is calculated based on the gain quantization index 66.
Decode 9, 80. In the pulse sound source, based on the sound source information codebook index 64, the pulse information codebook 7
The information 72 of the pulse interval and the head pulse position is read from 1, and the pulse generator 73 decodes the pulse sound source vector (pulse train) 74. In the noise source, the noise source codebook 75 outputs the noise source vector 76 in which the basis vectors are binary-weighted linearly combined based on the source information codebook index 64. Next, the used sound source selection unit 70 selects the used sound source based on the used sound source index 65 and outputs the sound source vector 77.

The second stage of the decoding process is the generation of the driving sound source. Long-term prediction lag index 6 from the adaptation codebook 68
The long-term predicted vector 69 and the sound source vector 77 read corresponding to 3 are multiplied by gains 79 and 80, and added to generate the driving sound source 84. Driving sound source 8
4 is input to the synthesis filter 85 and is also used for updating the state of the adaptive codebook 68.

The final stage of the decoding process is speech synthesis. The synthesis filter 85 uses the decoded short-term prediction parameter 82 as a filter coefficient, and inputs the driving sound source 84 to synthesize and output the digital synthesized speech 86.
Furthermore, in order to enhance the subjective sound quality, the synthesis filter 85
Output 86 is passed through a post filter 87 to obtain the final digital synthesized voice 88 as its output. This is continuously sent to the DA converter via the buffer memory and converted into analog synthesized voice.

The operations from voice input to encoding, decoding and voice output according to the embodiment of the present invention have been described above. In the above description, the frame energy (power) of voice is not particularly referred to. This is because the frame energy is reflected in the gain of the driving sound source, but considering the quantization of the gain, it is advantageous to normalize with the frame energy in advance in order to suppress the dynamic range of the gain. Since the frame energy can be easily obtained when calculating the linear prediction parameter, the frame energy is separately quantized and the index is transmitted. Also, in the case of long-term prediction analysis, the adaptive codebook is interpolated or the long-term prediction evaluation function is interpolated, and the fractional pitch is used to perform long-term prediction in sub-sample units. The voice quality can be improved. In the long-term prediction, the adaptive codebook and the long-term prediction vector are expanded / contracted in a waveform to cope with the temporal change of the periodic component, thereby improving the quality of the coded speech. When performing these processes, it is necessary to send information on the presence or absence of the fractional pitch or waveform expansion / contraction process to the decoding unit.

An example of bit allocation in this case is shown below.

Sampling frequency is 8 kHz, frame length is 4
0 ms (320 samples), subframe length 10 ms
(80 samples). It is assumed that the frame energy and the linear prediction parameter are updated in frame units, and the other parameters are updated in subframe units. It should be noted that it is effective to improve the quality of synthesized speech by interpolating the frame energy and the linear prediction parameter in subframe units. If the quantization is 27-bit multi-stage vector quantization, the quantization index of the linear prediction parameter is 27 bits. Frame energy is scalar quantized with 5 bits. Therefore, the number of transmission bits per frame is 3
It is 2 bits.

The sub-frame unit parameter has a long-term prediction lag index of 7 bits, which means that the range of the long-term prediction lag is 19 samples (421 Hz) to 146 samples (5
5 Hz). If two types of sound sources are used, the sound source switching index is 1 bit. The codebook size of the pulse and noise information codebook is 10 bits (1010
Code vector), the code vector index is 10 bits. The gain is expressed by 8 bits by converting the one for the long-term prediction vector and the one for the statistical code vector into different parameters, and then performing vector quantization. Therefore, the number of transmission bits per subframe is 25 bits. As a result, the total bit rate is 3300bp
s.

As described above, in the embodiment of the present invention, the reproducibility of the periodic component is improved and the quality is improved even if the bit rate is reduced. In addition, by searching the sound source codebook using a combination of impulse responses with truncated orders,
The processing amount can be reduced as compared with the conventional CELP.

[0105]

According to the present invention, the reproducibility of the periodic component, which is a problem when the CELP coding method is reduced to a low bit rate, is improved, and since it is used in combination with a noise source, it is 4 k.
It is possible to provide a voice coder having a good voice quality even at a bit rate of bps or less. Further, since the search processing amount of the pulse sound source can be reduced, it is possible to provide a speech encoder with a low processing amount.

[Brief description of drawings]

FIG. 1 is a block diagram of an encoding unit according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a decoding unit according to the first embodiment of this invention.

FIG. 3 is an explanatory diagram of the principle of a conventional CELP encoder.

FIG. 4 is an explanatory diagram of the principle of a conventional CELP decoder.

FIG. 5 is a correspondence diagram between a voice waveform and a residual waveform.

FIG. 6 shows an example of a pulse sound source.

FIG. 7 shows the structure of a pulse information codebook.

FIG. 8 is an explanatory diagram of the principle of pulse source vector generation.

FIG. 9 is a comparison of impulse response waveforms with and without weighting.

FIG. 10 is an explanatory diagram of a partial cross-correlation calculation method.

FIG. 11 is an explanatory diagram of a simplified convolution operation.

[Explanation of symbols]

12 ... Linear prediction analysis unit, 14 ... Linear prediction parameter quantization unit, 15, 81 ... Linear prediction parameter quantization codebook, 19 ... Auditory weighting filter, 21 ... Weighting synthesis filter, 25 ... Adaptive codebook search unit, 26, 6
8 ... Adaptive codebook, 31 ... Acoustic classification section, 34 ... Search codebook search section, 35, 70 ... Used sound source selection section, 3
6 ... Pulse sound source search unit, 40, 73 ... Pulse generation unit, 3
8, 71 ... Pulse information codebook, 37 ... Noise source search section, 42, 75 ... Noise information codebook, 51 ... Gain optimization / quantization section, 59, 78 ... Gain codebook, 8
5 ... Synthesis filter, 87 ... Post filter.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Atsushi Ishikawa 5-20-1 Kamimizuhoncho, Kodaira-shi, Tokyo Inside Hitachi Semiconductor Design Development Center

Claims (8)

[Claims] 1. An input voice signal is divided into frames of a predetermined time length, a spectrum parameter indicating a spectrum envelope of the voice signal is obtained and output, and the frame is divided into subframes of a predetermined time length. Then, a long-term prediction parameter is obtained from the past sound source so as to minimize the error from the speech signal and output, and an optimum code vector is selected from the codebook prepared in advance as a driving sound source for each subframe. In the CELP speech coding method, the codebook includes a plurality of codebooks, at least one of the plurality of codebooks is a noise component, and at least one is a pulse component having a constant amplitude and equally spaced. A speech coding method characterized by the following. 2. The code vector is selected by selecting a codebook to be searched from the plurality of codebooks according to a result of acoustic analysis of the voice signal, and selecting a codebook to be searched. The speech coding method according to claim 1, wherein the speech coding method is performed by selecting an optimum codebook from the code vectors. 3. The selection of the search codebook is characterized in that the entire codebook of the plurality of codebooks or a part of each codebook is set as a search target codebook. 2. The voice encoding method according to 2. 4. The codebook of the pulse component stores information on the position of the first pulse and the pulse interval as the information of the pulse train, and the code vector of the pulse train is generated from the information. A speech coding method according to any one of claims 1 to 3. 5. The speech coding method according to claim 1, wherein the optimum code vector of the pulse component is selected by a full search of the code book. 6. The speech coding method according to claim 1, wherein the interval between the pulse trains is a range that substantially covers a variation range of a pitch period of human utterance. 7. The head pulse position in the sub-frame of the pulse train can be taken from the head to the last point of the sub-frame regardless of the pulse interval. Speech coding method. 8. The code vector of the pulse component is selected based on a combination of impulse responses of a weighting synthesis filter whose length is truncated to a value equal to or less than a minimum value of the interval of the pulse train. The speech coding method according to any one of claims 1 to 7.