JPH05273999A - Voice encoding method - Google Patents

Abstract

PURPOSE:To provide the voice encoding method which can obtain high-quality synthetic voice even at a low bit rate lower than 4kbps. CONSTITUTION:This method is provided with a pulse component adaptive code book 34, noise component adaptive code book 31, pulse generator 42, first noise code book 46, second noise code book 50 and sound source selector 54 at the voice encoding part, and the sound source code book is switched by weighted error evaluation. Thus. the reproduciveness of the cyclic component of voice is improved, and high-quality voice can be obtained even at the 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 more particularly to a speech coding method capable of reducing the bit rate to 4 kbps or less.

[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 recently 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 BSAtal: "Code-exci").
ted linear prediction (CELP) ", Proc. ICASSP 85 (19
85.3)) 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 Gerson and MA).
Jasiuk: "Vector sum excited linear prediction (VS
ELP) speech coding at 8kbps ", Proc. ICASSP 90 (199
0.4)) is excellent in processing amount, memory capacity, and bit error resistance.

On the other hand, the digitization of wireless communication has begun in earnest,
From the viewpoint of effective use of frequencies, it is desired to develop a voice encoding system with a lower bit rate (4 kbps or less). If CELP or VSELP is simply made to have a low bit rate, quality deterioration becomes large and there is a limit.
It is considered that this is because the reproducibility of the periodic component in the voice is lowered by lowering the bit rate. Therefore,
A method that employs a sound source that enhances the reproducibility of the periodic component has been proposed.

As such a system, "MPC-CEL" which uses multi-pulse for voiced sound and CELP for unvoiced sound is used.
P "method (Ozawa, Kumagai:" 3.2 kb / s speech coding method using multi-pulse and CELP ", IEICE Spring National Convention (1990.3)), and phase and amplitude control for voiced sounds. Single pulse, CELP for unvoiced sounds
"SPE-CELP" method (W. Granzow and
BSAtal: "High-quality digital speech at 4 kb /
s ", Proc. GLOBECOM 90 (1990.12)), etc.
"Spindle adaptive VXC" method using pulse and noise orthogonalized to it as a sound source (Tanaka et al. 3: "CELP coding by multiple vector synthesis", Acoustical Society of Japan Proceedings 1-3-5 (1989.10) ), Or “pulse / noise selection type CEL” that is used by switching 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)) has also been proposed.

[0005]

The above-mentioned proposed system has the following problems. When switching between essentially different encoding methods (for example, multi-pulse and CELP), the sound quality tends to be unnatural such as a change in tone color. Also, when switching between pulse and noise, the sound quality when using pulses tends to be pulsive or buzzer-like. Furthermore, in the method that uses both pulse and noise,
There is a problem that the bit rate cannot be reduced sufficiently.

It is an object of the present invention to provide an encoding method in which the reproducibility of the periodic component of voice is high and the change of the tone color is not noticeable even if the bit rate is reduced.

[0007]

In order to achieve the above object, the present invention has the following means. It has (1) a pulse generator, (2) a first noise codebook with a small codebook size, (3) a second noise codebook with a large codebook size, and (4) a weighting error evaluator.
Further, another embodiment of the present invention has (5) acoustic classifier, (6) pitch extractor, (7) pulse component adaptive codebook, and (8) noise component adaptive codebook.

[0008]

The operation of the typical structure of the present invention will be described.
The voice input to the encoder is first divided into frames and subframes. The short-term prediction analysis unit extracts and quantizes the spectrum parameter (short-term prediction coefficient) for each frame. Next, the input speech is perceptually weighted in preparation for evaluating perceptual weighting errors. Also, input the zero signal to the weighting synthesis filter,
The quiescent response is determined and subtracted from the weighted input signal. This is to remove past effects that depend on the internal state of the synthesis filter.

Next, the long-term prediction analysis unit obtains the optimum long-term prediction lag and gain from the adaptive codebook in subframe units. The adaptive codebook is divided into a pulse component and a noise component, and the long-term prediction lag is searched for the long-term prediction filter component obtained by the weighted sum with the optimum gain.

In the pulse generator, the pulse position is 1 with the long-term prediction lag obtained by the long-term prediction analyzer as the pulse interval.
It is generated by shifting each sample and weighted by convolving the impulse response of the weighting synthesis filter. After orthogonalizing these with respect to the long-term prediction vector, a pulse source located at a position where the weighting error is minimized is searched, and the position and the gain are determined.

The first codebook search section weights the code vectors in the first noise codebook in the same manner as the pulse sound source. The code and gain of the code vector that minimizes the weighting error are determined by orthogonalizing the long-term prediction vector and the pulse source vector.

The search for the second noise codebook can be performed concurrently with the search for the pulse sound source and the search for the first noise codebook. In the second codebook search unit, the code vector in the second noise codebook is
Weighting is performed in the same manner as the pulse sound source. The code and gain of the code vector that is orthogonalized to the long-term prediction vector and minimizes the weighting error are determined.

The selector evaluates the weighting error between the case where the pulse sound source and the first noise sound source are used and the case where only the second noise sound source is used, and the smaller error is taken as the final sound source. select.

The gain quantizer simultaneously optimizes and quantizes the gains of the sound sources selected by the selector.

The spectrum parameter, the quantized code of gain, the long-term prediction lag, the position of the selected pulse sound source, and the index of the noise code vector 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 having the short-term prediction coefficient as a filter coefficient, whereby decoded speech is obtained.

In another configuration of the present invention, it is also possible to control the weighting error evaluation in the selector by performing acoustic classification. Furthermore, it is also possible to perform pitch extraction and use the extracted pitch period as the pulse interval of the pulse sound source.

[0018]

Embodiments of the present invention will be described below with reference to the drawings.

The present invention is directed to code driven speech coding (CEL).
Since it is based on the P) method, the outline of the principle of the CELP method will be described first. In CELP coding, a long-term prediction vector that is an output of an adaptive codebook as a component representing the periodicity of a sound source, and components other than the periodicity (randomness
Alternatively, the sum of weights obtained by multiplying the code vectors, which are the output of the noise codebook (also referred to as statistical codebook) as the noise characteristics, by the respective gains and adding them is used as the 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 the 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 is input to the weighting synthesis filter to obtain the weighted synthesis speech. The input speech is also obtained through the weighting filter to obtain the weighted input speech, and the difference from the weighted synthesized speech is obtained to obtain the weighted error waveform. For the weighted error waveform, the sum of squares is calculated over the error evaluation section, and the weighted squared error is obtained. Since the driving sound source is the weighted sum of the long-term predicted vector and the noise 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 and the code vector index are sequentially changed to calculate the weighted squared error, and the one with the smallest weighting error may be selected. Such a driving sound source determination method is called a "synthesis analysis" method. If the above procedure is faithfully performed, that is, if the long-term prediction lag and the index of the noise code vector are optimized at the same time while evaluating the weighting error, a huge amount of processing is required. Is used.

FIG. 1 shows a block diagram of an encoding unit according to the first embodiment of the present invention, and FIG. 2 shows a block diagram of a decoding unit. The outline of the operation of the first embodiment will be described below.

The voice encoding unit receives the digital voice signal 11 A / D converted at a predetermined sampling frequency (usually 8 kHz). The sound classifier 12 classifies the input voice into a plurality of categories, such as vowel characteristics and frictional characteristics, based on the acoustic characteristics of the input voice. The sound classification result is output as the sound classification flag 13.

The short-term prediction analyzer (LPC analyzer) 17 reads out the speech data 11 of the analysis frame length and outputs the short-term prediction coefficient 18. The frame length is, for example, 40 ms (3
20 samples).

The short-term prediction coefficient 18 is quantized by the short-term prediction coefficient quantizer 19. The quantized code is output as the transmission parameter as the short-term prediction coefficient quantization index 21. In addition, the quantized value 20 of the short-term prediction coefficient is referred to in the subsequent processing.

Further, the input voice is weighted by the auditory weighting device 22 to obtain a weighted voice 23. On the other hand, a signal (zero input) 25 having a value of 0 for the frame length is input to the weighting synthesis filter 24, and a zero input response 26 is obtained. This is subtracted from the weighted input voice 23 to obtain the weighted input voice 27 in which the influence of the past internal state of the weighting synthesis filter is removed.

Since the long-term predictive analysis is performed by searching the adaptive codebook for each subframe, it will be referred to as adaptive codebook search hereinafter. Here, the subframe length is, for example, about 8 ms (64 samples). The present invention has two adaptive codebooks for the pulse component and the noise component, and the adaptive codebook searcher 3 is also shown in the drawing.
1, 34. As described later, the states of the two adaptive codebooks are combined, the long-term prediction lag 37 that is a parameter representing the periodicity of the speech is extracted, and the long-term prediction lag index 38 and the long-term prediction vector 41 are output.

In the pulse generator 42, the long-term prediction lag 37 obtained by the long-term prediction analysis is used as pulse intervals, the pulse positions are shifted by one sample, and the pulses are generated by weighting by convolution of the impulse response of the weighting synthesis filter. After orthogonalizing these with respect to the long-term prediction vector 41,
The pulse sound source located at the position where the weighting error is minimized is searched, and the position 44 and the gain 43 are determined. The weighted pulse excitation vector 45 of the selected pulse component is output to the first noise codebook search unit 46.

In the first codebook search section 46,
The code vectors in the noise codebook of 1 are weighted in the same manner as the pulse source. The code 48 and the gain 47 of the code vector that minimizes the weighting error by orthogonalizing the long-term prediction vector 41 and the pulse sound source vector 45.
To decide.

The search of the second noise codebook can be executed in parallel with the search of the pulse sound source and the search of the first noise codebook. Second codebook search unit 56
Then, the code vectors in the second noise codebook are weighted in the same manner as the pulse source. The code 52 and the gain 51 of the code vector that is orthogonalized to the long-term prediction vector 41 and minimizes the weighting error are determined.

The selector 54 evaluates the weighting error between the case where the pulse sound source and the first noise sound source are used and the case where only the second noise sound source is used, and the smaller error is the final sound source. Is output as a sound source selection flag 55. Here, the weighting error is corrected by the sound classification flag 13, and the sound source that improves the subjective quality is preferentially selected.

The gain codebook searcher 56 simultaneously optimizes the gains of the sound sources selected by the selector 54 by searching the gain codebook, and outputs the quantized code 57 at that time.

Quantization codes 21, 57 of the short-term prediction coefficient and gain, the index 38 of the long-term prediction lag, the position 44 of the selected pulse sound source, the noise code vector indexes 48, 52, and the sound source obtained as described above. The selection flag 55 is transmitted as a transmission parameter to the decoder.

In the speech decoding unit of FIG. 2, the codebooks 38 ', 44', 48 'and 52' are used to extract the codevectors 62 from the codebooks 61, 65, 77 and 81.
66, 78, 82 are read out and the pulse generator 7
3 produces a pulsed sound source 78. The gains 63, 67, 75, 79, 83 are reproduced from the gain codebook using the gain codebook index 57 '. The driving sound source vector 89 is generated by multiplying each gain of each code vector. However, one of the pulse sound source and the first noise sound source or the second noise sound source is selected by the switch 87 based on the sound source switching flag 55 ′.

The driving sound source 89 is converted into a short-term prediction coefficient 21 '.
By inputting it to the synthesis filter 93 having a filter coefficient of, a synthesized voice 94 is obtained. Finally, for the purpose of improving subjective sound quality, the synthesized speech 94 is input to the adaptive post filter 95, and the final decoded speech 96 is obtained.

The decoded voice (digital signal) 96 is DA converted, converted into analog voice, and output.

Now that the outline has been described, the detailed functions of the main parts of the first embodiment will be described.

The acoustic classifier 12 calculates physical parameters from the voice data 11 of frame length or subframe length, and classifies the voice of the section into a plurality of categories by logically judging the parameter values. is there. The sound classification method itself is a known technique, for example, Ozawa: “4.8 kb / s multi-pulse speech coding method using various sound sources”, Proceedings of Acoustical Society of Japan (198).
An example is disclosed in 9.3). An example of the configuration as an acoustic classifier is (Chuken) reception number 31920044.
8 patents. As physical parameters,
For example, energy, energy change rate, maximum correlation value,
Predicted gain, logarithmic cross-sectional area ratio, etc. are used. The categories of voice are classified into vowels, nasal sounds, plosive / transient, frictional properties, and vowels / nasal sounds, rising and falling. Acoustic classification is performed in frame units or subframe units. For example, when calculating the energy change rate in frame units, the difference between the frame energy of the previous frame and the frame energy of the current frame or the energy of each subframe is calculated. The change may be calculated. Further, when the calculation is performed in units of subframes, the energy difference between adjacent subframes or the subframe is further divided into the first half and the second half, and the difference in energy of each may be detected.

The short-term predictive analyzer (LPC analyzer) 17 is
The short-term prediction coefficient 18 representing the speech spectrum envelope is extracted from the speech data 11 for each frame. Short-term prediction coefficient 1
8 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 method of deriving the reflection coefficient is
In addition to the above, the FLAT algorithm (established by the Radio System Development Center, "Digital Car Telephone System Standard RCR STD-27" (hereinafter "RCR Standard")
Abbreviated) and the LeRoux method (Saito,
Nakata, described in the above-mentioned book) has been 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 predictive coefficient 18 is converted into an LSP parameter, which is then subjected to a two-stage vector quantization by a quantizer 19 and converted into a quantized value 20. LSP
It is known that the parameter has a better quantization characteristic than that of directly quantizing a linear prediction coefficient (spectrum 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, vector quantization, or vector / scalar quantization may be used. The quantization index 21 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 unit 22 first weights the input voice 11 to obtain a weighted voice 23. The weighting filter is composed of the quantized value 20 of the short-term prediction coefficient (or an equivalent parameter), and its specific form is as follows.

[0043]

[Equation 1]

Here, αi 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 23 in advance. That is, data having a value of 0 corresponding to the frame length in the weighting synthesis filter 24 (zero input 25)
, The zero input response 26 is calculated, and the weighted speech 23
Weighted speech 27 subtracted from
To get The transfer function of the weighting synthesis filter 24 used here is as follows.

[0046]

[Equation 2]

The synthesizing filter 24 is different from the synthesizing filter on the decoding side in that it includes the weighting parameter λ.

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, a case where the noise codebook is sequentially searched will be described. That is, assuming that the output of the noise codebook is 0, the optimum long-term prediction vector 41 is determined.

Since the present invention has two adaptive codebooks for the pulse component and the noise component, the codebook search is as follows. The weighted squared error is defined by the following equation.

[0050]

[Equation 3]

Where b _LP (n): pulse component codebook a for lag L
Output of _P (n) b _LN (n): Noise component codebook a _N (n) for lag L
Output _{b 'LP (n): b} LP (n) of the weighted synthesized speech _{b' LN (n): b} LN (n) of the weighted synthesized speech beta _P: gain of pulse component beta _N: gain of the noise component p ( n): Weighted input speech with the past influence removed. However, the weighted synthesis is realized by convolving the output of the codebook with the impulse response of the weighted synthesis filter. The synthesized output thus obtained does not depend on the past states of the synthesis filter and is therefore called the zero-state response. By partially differentiating (Equation 3) with β _P and β _N , the optimum gain is

[0052]

[Equation 4]

The squared error at this time is

[0054]

[Equation 5]

It becomes However,

[0056]

[Equation 6]

It is Therefore, the optimum lag L may be a lag that maximizes the right-hand side two terms of (Equation 5). However, only those for which β _P and β _{N are} positive are targeted. The long-term prediction vector 41 becomes b ′ _L (n) = β _P b ′ _LP (n) + β _N b ′ _LN (n). However, L is the optimum lag.

Next, generation of a pulse sound source in the pulse generator 42 will be described. The pulse generator 42 generates a pulse train c _P (n) having an optimum lag L as a pulse interval. The pulse train can be uniquely determined by using the position of the leading pulse in the subframe as an index. As shown in FIG. 3, the relationship between the pulses in the sub-frame depends on the relationship between the lag L and the sub-frame length N.
It can be divided into two types.

Type 1 is a case where L _min ≤L≤N-1 and there are a plurality of pulses in a subframe, and there are L arrangements. Type 2 is N ≦ L ≦ L
_In case of _max , there is only one pulse in a subframe. At this time, there are N arrangements. However, L _min and L
_max is the minimum value and the maximum value of the search range of lag. For example, if L _min = 20, L _max = 146, N = 64,
Depending on the value of L, all the arrangements can be represented by 5 bits or 6 bits.

Next, the search for the pulse sound source will be described.
Generate the weighted synthesized speech of the generated pulse sound source c _P (n) as f _P (n)
Then, f _P (n) is orthogonalized to b ′ _L (n), and f ′ _P (n)
about

[0061]

[Equation 7]

Find f _P (n) that minimizes For the orthogonalization, the Gram-Schmidt orthogonalization method or the like is used.

Next, the first noise codebook search unit 46
The search for the noise source (1) in (1) will be described. Assuming that the total number of bits allocated to the pulse sound source and the noise sound source (1) is 10 bits, 5 bits or 4 bits can be allocated to the noise sound source depending on the value of the lag L. This is to search for a noise codebook whose codebook size is 5 bits or 4 bits.

Let f _N1 (n) be the weighted synthesized speech of the noise code vector c _N1 (n). 'When _{N1 (n), f' f} N1 and b _'L and f (n)' _P the vector is orthogonal to the (n) (n) f For _N1 (n)

[0065]

[Equation 8]

Find f _N1 (n) that minimizes

Next, the second noise codebook search unit 50
The search for the noise source (2) in (1) will be described. A 10-bit noise codebook equal to the total number of bits assigned to the pulse sound source and the noise sound source (1) will be searched.

Let f _N2 (n) be the weighted synthesized speech of the noise code vector c _N2 (n). When f _N2 and b _'L vector was orthogonal to the (n) f' (n) and _N2 (n), the f _'N2 (n)

[0069]

[Equation 9]

Find f _N2 (n) that minimizes The noise codebook may have a VSELP type structure. By doing so, it is possible to significantly reduce the processing amount as compared with the normal CELP. VSELP
Are described in detail in the RCR standard mentioned above.

Next, selection of a sound source by the selector 54 will be described. Basically, the weighting error when using the pulse sound source and the noise sound source (1) and the noise sound source (2)
The weighting error in the case of using only is compared, and the sound source with the smaller error is selected. The output of the selector 54 is a sound source selection flag 55.

The weighting error when the pulse sound source and the noise sound source (1) are used is as follows.

[0073]

[Equation 10]

Similarly, the weighting error when only the noise source (2) is used is as follows.

[0075]

[Equation 11]

Here, β _P , β _N , γ _P , γ _N1 and γ _N2 are gains. In reality, the gain codebook searcher 56 searches the gain codebook, and the quantized gains obtained as a result are used to evaluate (Equation 10) and (Equation 11).

Here, the contribution of the sound classification result will be briefly described. When the weighting errors of (Equation 10) and (Equation 11) are evaluated, it is possible that the selected sound source is frequently switched due to the effect of quantization of gain or the like. In such a case, the subjective quality may not necessarily be good. For example, in the stationary part of vowels, it is better to fix it to one side even if the weighting error is somewhat bad. Therefore, the weighting error of (Equation 10) or (Equation 11) is biased based on the sound classification result.

Next, updating of the adaptive codebook will be described. Here, first, the same driving sound source as in the decoder is created. When the pulse sound source and the noise sound source (1) are selected, the driving sound source is as follows.

Ex ₁ (n) = ex _P1 (n) + ex _N1 (n) where ex _P1 (n) = β _P b _LP (n) + γ _P c _P (n) ex _N1 (n) = β _N b _LN (n) + γ _N1 c _N1 (n). On the other hand, when the noise source (2) is selected, the driving source is ex ₂ (n) = ex _P2 (n) + ex _N2 (n), where ex _P2 (n) = β _P b _LP (n) ex _N2 (n) = β _N b _LN (n) + γ _N2 c _N2 (n).

Updating the adaptive codebook for pulse components is
It is as follows. a _P (n) = a _P (n + L): −L _max ≦ n <−N a _P (n) = ex _Pi (n + L): −N ≦ n <0 However, when i = 1, pulse source and noise When the sound source (1) is selected When i = 2, the noise source (2) is selected.

On the other hand, the noise component adaptive codebook is updated as follows. a _N (n) = a _N (n + L): −L _max ≦ n <−N a _N (n) = ex _Ni (n + L): −N ≦ n <0 However, when i = 1, pulse source and noise When the sound source (1) is selected When i = 2, the noise source (2) is selected.

The generation of the driving excitation and the updating of the adaptive codebook are exactly the same in the encoder and the decoder.

Next, returning to FIG. 2, the decoding unit of this embodiment will be described. As described in the overview, the decoder obtains the driving sound source by multiplying the code vector obtained by searching the codebook based on each index or the pulsed sound source with the decoded gain and adding it. .. By inputting this into a synthesis filter having a line station prediction coefficient as a filter coefficient, a synthetic sound is obtained.

The actual sound source is switched based on the sound source selection flag. For details, refer to the description of the encoder.

According to this embodiment, the CE having a low bit rate is used.
Also in the LP encoder, the reproducibility of the periodic component is improved and the quality can be improved.

Next, a second embodiment of the present invention will be described. FIG. 4 shows a block diagram of the speech coder of this embodiment. This embodiment shows a case where the acoustic classifier is not used, and the selector 54 selects a sound source only by comparing the weighting errors of (Equation 10) and (Equation 11). The speech decoder is the same as that of the first embodiment. Since sound classification is not performed in this embodiment, the processing amount can be reduced compared to the first embodiment.

Next, a third embodiment of the present invention will be described. FIG. 5 shows a block diagram of the speech coder of this embodiment. In the present embodiment, the pitch extractor 14 is provided to extract the pitch period of the input speech 11 and output the value 15 and the code 16. The pitch period 15 defines the pulse interval generated by the pulse generator 42. That is, in the first embodiment, the pulse interval is determined to match the long-term prediction lag, but in the present embodiment, the pulse interval matches the pitch period.

The code 16 of the pitch period is transmitted to the decoder as a transmission parameter.

FIG. 6 shows a block diagram of the decoder of this embodiment. The difference from the decoder of the first embodiment is that the pulse interval of the pulse sound source is determined from the code of the pitch period. In the present embodiment, the voice quality is further improved by generating the pulse reflecting the physical phenomenon of the voice called the pitch period.

Next, a fourth embodiment of the present invention will be described. FIG. 7 shows a block diagram of the speech coder of this embodiment. In the present embodiment, the pitch extractor 14 is provided as in the third embodiment, but the pitch period 15 is set to the pulse generator 4.
The feature is that it is not directly reflected in the pulse interval in 2. In this example, pitch period 15 affects the choice of long-term predicted lag. That is, in the first embodiment, the long-term predicted lag is determined to minimize E of (Equation 5), but in the present embodiment, when the searched lag L is close to the pitch period 15, E of E is calculated. Bias to lower the value at a constant rate. By doing so, the long-term prediction lag, which is normally easy to randomly take a value that is an integral multiple of the pitch period, has a high rate of matching with the pitch period, and the continuity of the long-term prediction lag is improved.

Further, since the pulse interval in the pulse generator 42 matches the long-term prediction lag as in the first embodiment, it is not necessary to transmit the pitch period index. The audio decoder is the same as the decoder of the first embodiment.

According to this embodiment, the continuity of the long-term prediction lag is improved, and at the same time, the pulse interval of the pulse sound source also reflects the physical characteristics of the voice, so that the voice quality can be improved.

Next, a fifth embodiment of the present invention will be described. FIG. 8 is a block diagram of the speech coder of this embodiment. The difference from the previous embodiments is that the adaptive codebook 3
9 is not separated for the pulse component and the noise component, and there is only one. Correspondingly, there is only one speech decoder and one adaptive codebook 69 shown in FIG. By doing so, it is possible to search for the long-term prediction lag and reduce the update processing amount. This embodiment is effective when it is desired to realize a low throughput.

[0094]

According to the present invention, the reproducibility of the periodic component which is a problem when the CELP encoder is made to have a low bit rate is improved, so that the speech encoder having a good voice quality even at a bit rate of 4 kbps or less. Can be provided.

[0095]

[Brief description of drawings]

FIG. 1 is a block diagram of an encoding unit according to a first embodiment of this 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 pulse sound source generation.

FIG. 4 is a block diagram of an encoding unit according to a second embodiment of the present invention.

FIG. 5 is a block diagram of an encoding unit according to a third embodiment of the present invention.

FIG. 6 is a block diagram of a decoding unit according to a third embodiment of the present invention.

FIG. 7 is a block diagram of an encoding unit according to a fourth embodiment of the present invention.

FIG. 8 is a block diagram of an encoding unit according to a fifth embodiment of the present invention.

FIG. 9 is a block diagram of a decoding unit according to a fifth embodiment of the present invention.

[Explanation of symbols]

 12 ... Acoustic classifier, 14 ... Pitch extractor, 17 ... Auditory weighter, 24 ... Weighting synthesis filter, 31, 61 ... Noise component adaptive codebook searcher, 34, 65 ... Pulse component adaptive codebook searcher, 39 ... Adaptive codebook searcher, 42 ... Pulse generator, 46 ... First noise codebook searcher, 50 ... Second noise codebook searcher, 54 ... Sound source selector, 56 ... Gain codebook searcher, 58 ... Multiplexing device, 59 ... Transmission line, 60 ... Demultiplexing device, 69 ... Adaptive codebook, 73 ... Pulse generator, 77 ... First noise codebook, 81 ... Second noise codebook, 87 ... Sound source switching device , 93 ... Synthesis filter, 95 ... Adaptive post filter, 97 ... Gain codebook.

Claims (7)

[Claims] 1. A code-driven linear prediction (CELP) coding method in which a plurality of sound sources are switched and used, wherein at least one of the sound sources is a combination of a pulse component and a noise component, and a pulse of the pulse component. The position and the gain are set independently of the noise vector of the noise component. 2. The speech coding method according to claim 1, wherein the switching of the sound sources is determined by evaluating a weighting error in a coding target section. 3. The speech coding method comprises an acoustic classifier for classifying an input speech based on acoustic characteristics of the speech.
The speech coding method according to claim 2, wherein the classification result of the acoustic classifier is reflected in the evaluation of the weighting error. 4. The pulse interval of the pulse component is made to coincide with a long-term prediction lag within the coding target section, and the position of the first pulse of the pulse component is set at a position from the beginning of the coding target section. The audio encoding method according to claim 1, wherein the audio encoding method is defined and used as an index. 5. The voice encoding method comprises a pitch extractor for extracting a pitch period of an input voice, wherein a pulse interval of the pulse component is matched with a pitch period extracted by the pitch extractor, 5. The speech coding method according to claim 1, wherein the position of the first pulse of the pulse component is defined as a position from the beginning of the coding target section and is used as an index. .. 6. The speech coding method according to claim 4, further comprising a pitch extractor for extracting the pitch period of the input speech, and the output of the pitch extractor is reflected in the determination of the long-term prediction lag. The audio encoding method described in. 7. The speech encoding method comprises an adaptive codebook for pulse components and an adaptive codebook for noise components,
2. The search for the long-term prediction lag is performed for a long-term prediction filter component synthesized by optimizing the gain of the adaptive codebook for pulse components and the gain of the adaptive codebook for noise components. The audio encoding method according to claim 6.