JP3192999B2 - Voice coding method and voice coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a low bit rate speech coding / decoding method, and more particularly, to a coding / decoding method of an excitation signal for performing phase adaptation of a noise excitation signal.

[0002]

2. Description of the Related Art The CELP system is known as a typical system for low bit rate speech coding. The CELP method is a method for expressing a speech signal by coding spectral envelope information and a sound source signal. As an encoding method of a sound source signal for further improving sound quality in the CELP method, attention has been paid to phase adaptation of a noise sound source signal. Have been. This phase adaptation method is described in, for example, Mino et al., “Phase Adaptive PSI-CEL”.
Examination of P speech coding ", IEICE Technical Report SP
94-96, p. 37-44, February 1995 (Literature 1). A speech coding system using the conventional phase adaptation method described in this reference 1 will be described with reference to FIG.

In FIG. 4, an input audio signal from an audio input terminal is first analyzed by a linear prediction analysis unit 502 to extract a linear prediction coefficient (LPC coefficient). The extracted linear prediction coefficients are quantized by the synthesis filter information quantization unit 503, and the quantized linear prediction coefficients, that is, the code A indicating the synthesis filter information indicating the characteristics of the synthesis filter unit 504 is transmitted to the multiplexing unit 507. Is output. The quantized linear prediction coefficients are used as filter coefficients of the synthesis filter unit 504.

An adaptive codebook 501 stores past encoded excitation signals, and stores past excitation signals in a pitch cycle L
To generate an adaptive code vector. The search for the pitch period L is performed by adding the adaptive code vector candidate corresponding to the pitch period candidate to the gain multiplication unit 508 and the addition unit 50.
9, the distortion of the waveform when synthesized as speech by the synthesis filter 504 is calculated by the auditory weighted distortion calculator 505 via the subtractor 522 using the auditory weighted distortion scale, and the distortion of the synthesized auditory weighted waveform is calculated. This is performed by searching the code selection unit 506 for a smaller pitch period. The code representing the pitch period L searched in this way is output to the multiplexing unit 507.

[0005] The phase search section 512 obtains a phase shift amount using the adaptive code vector from the adaptive codebook 501 and the pitch period L searched by the code selection section 506 as described later. Output to

[0007] The noise codebook 520 stores a plurality of noise vectors, and outputs a noise vector corresponding to the noise code C searched by the code selection unit 506 to the phase adaptation unit 519. Phase adaptation section 519 cuts out a noise vector output from noise codebook 520 from a position corresponding to the amount of phase shift provided from phase search section 512. Pitch periodicizing section 518 gives the extracted noise vector pitch periodicity, and outputs this as a noise code vector.

In searching for the noise code C, a noise code vector candidate corresponding to the noise code candidate is synthesized as speech by a synthesis filter 504 via a phase adaptation section 519, a pitch periodization section 518, a multiplication section 510 and an addition section 509. The distortion of the waveform at that time is calculated by the auditory weighted distortion calculator 505 via the subtractor 522 by the auditory weighted distortion calculator, and the noise selector that reduces the distortion of the auditory weighted composite waveform is reduced by the code selector 506. This is done by searching. The noise code C searched in this way is output to the multiplexing section 507.

The gain codebook 521 stores a gain G0 applied to an adaptive code vector and a candidate G1 applied to a noise code vector. The search for the gain code is
Waveforms when the excitation filter candidates generated by adding the adaptive code vector and the noise code vector obtained by multiplying the gain G0 and G1 candidates by the gain multiplication units 508 and 510 by the addition unit 509 are synthesized by the synthesis filter 504. Of the auditory weighting through the subtractor 522
The calculation is performed by using a perceptually weighted distortion scale in step S05, and the code selecting unit 506 searches for a gain code in which the distortion of the perceptually weighted synthesized waveform becomes smaller. The gain code G searched in this way is output to multiplexing section 507.

The phase search section 512 includes a pulse train setting section 5
13 and a synthesized waveform obtained by synthesizing a pulse train (referred to as a pitch pulse train) at intervals of a pitch period L from the adaptive codebook 501 with a synthesis filter 514 having a transfer characteristic represented by 1 / Aq (z). The phase of the pitch pulse train that minimizes the error (waveform distortion) between the adaptive code vector and the synthesized waveform synthesized as speech by the synthesis filter 515 having the transfer characteristic represented by 1 / Aq (z) Phase candidate setting section 516 and optimal phase selection section 51
7, the amount of phase shift in the phase adaptation unit 519 is obtained.

Here, the output of the synthesis filters 514 and 515 of the transfer characteristic 1 / Aq (z) is a synthesized waveform, that is, the level of a voice waveform. Phase search,
That is, it can be said that the search for the phase shift amount for extracting the noise vector output from the noise codebook 320 is performed. However, in this method, a phase shift amount that is acoustically preferable is not always searched, so that when a low-rate coding such as about 4 kbit / s is performed,
There is a problem that encoding distortion of the excitation signal is easily perceived.

[0011]

As described above, in the conventional speech coding / decoding method, the phase search of the pitch pulse train, that is, the noise vector output from the noise codebook is cut out by evaluating the distortion of the speech waveform level. The amount of phase shift searched for is not always acoustically preferable because the amount of phase shift searched for is 4 kb.
In low-rate coding at about it / s, there is a problem that coding distortion of the excitation signal is easily perceived.
An object of the present invention is to provide a speech encoding / decoding method in which encoding distortion is hardly perceived even in low-rate encoding.

[0012]

In order to solve the above-mentioned problems, a speech encoding / decoding method according to the present invention comprises a pseudo-auditory weight synthesis filter for performing pseudo-aural weighting from synthesis filter information. A phase shift amount for extracting a noise vector output from the noise codebook is searched at a synthesized waveform level with a pseudo auditory weight by a pseudo auditory weight synthesis filter.

That is, the speech encoding method of the present invention comprises:
(a) generating an adaptive code vector by repeating a drive excitation signal output from an adaptive codebook storing past drive excitation signals at a pitch cycle, and (b) a waveform synthesized from a pulse train repeated at a pitch cycle. Phase shift amount search for searching a phase shift amount for extracting a noise vector output from a noise codebook storing a plurality of noise vectors so that an error between the waveform and a waveform synthesized from the adaptive code vector becomes smaller. (C) generating a noise code vector by periodicizing a noise vector cut out according to the amount of phase shift at a pitch period, and (d) according to synthesis filter information obtained from an analysis result of an input speech signal. By driving the first synthesis filter whose characteristics have been determined using the adaptive code vector and the noise code vector, (E) encoding synthesis filter information representing the characteristics of the first synthesis filter; and (f) a pitch period in which the distortion of the synthesized speech signal after auditory weighting becomes smaller. And searching for a noise vector.

The phase shift amount searching step includes a second step of synthesizing voice by performing pseudo auditory weighting.
, Based on the synthesis filter information, and a synthesized waveform synthesized from the pulse train by the second synthesis filter and a synthesized waveform synthesized from the adaptive code vector by the second synthesis filter. The phase shift amount is searched such that the error between the phase shift and the phase shift is smaller.

On the other hand, the speech decoding method of the present invention corresponding to this speech encoding method comprises the steps of (a) repeating a drive excitation signal output from an adaptive codebook storing past drive excitation signals at a pitch cycle. Generating an adaptive code vector; and (b) storing a plurality of noise vectors so that an error between a waveform synthesized from a pulse train repeated at a pitch period and a waveform synthesized from the adaptive code vector becomes smaller. A phase shift amount searching step of searching for a phase shift amount for cutting out a noise vector output from the codebook, and (c) a noise code vector by periodicizing a noise vector cut out according to the phase shift amount with a pitch period. Generating a first synthesis filter whose characteristics are determined in accordance with the synthesis filter information; Generating a reproduced audio signal by driving using the vector.

Then, in the phase shift amount searching step, a second synthesis filter (pseudo-auditory weight synthesis filter) for synthesizing speech by performing pseudo-auditory weighting from the synthesis filter information is configured. The phase shift amount is searched such that the error between the synthesized waveform synthesized from the pulse train by the synthesis filter of the above and the synthesized waveform synthesized from the adaptive code vector by the second synthesis filter becomes smaller.

As described above, according to the speech encoding / decoding method of the present invention, the phase for extracting the noise vector output from the noise codebook at the synthetic waveform level with pseudo auditory weights by the pseudo auditory weight synthesis filter. By searching for the shift amount, it is possible to search for a phase shift amount that is acoustically preferable. Therefore, even at a low rate of about 4 kbit / s, encoding / decoding of the excitation signal in which encoding distortion is hardly perceived can be realized.

In the speech encoding / decoding method according to the present invention, the impulse response of the pseudo-auditory weighting synthesis filter may be cut off at a finite number of predetermined samples, thereby substantially reducing the phase search performance. Without this, it is possible to greatly reduce the amount of calculation of filtering required for synthesis by the pseudo auditory weight synthesis filter.

[0019]

Embodiments of the present invention will be described below with reference to the drawings. [Speech Coding System] FIG. 1 is a block diagram showing a configuration of a speech coding system to which a speech coding method according to an embodiment of the present invention is applied.
1 shows an example in which phase adaptation is used for encoding a noise source in the system. In the CELP method, a vocal cord signal is made to correspond to a sound source signal as a model of a speech generation process, a spectrum envelope characteristic represented by a vocal tract is expressed by a synthesis filter, and the sound source signal is input to a synthesis filter, and the output of the synthesis filter is It is fundamental to represent an audio signal. At this time, the sound source signal is encoded so that the error between the audio signal input to the encoding system and the audio signal reproduced by the decoding system, that is, the error of the reproduced audio signal is perceptually reduced. There is a feature of the present invention.

The characteristics of the filter used in the phase search unit differ greatly between the speech coding system of the present embodiment and the conventional speech coding system shown in FIG. That is, in the phase search unit 512 in the conventional speech coding system,
While the phase search is performed by the distortion evaluation at the audio waveform level by the synthesis filters 514 and 515 in FIG. 4, the phase search unit 112 in the present embodiment sets the pseudo-auditory weight synthesis filter as shown in FIG. A transfer characteristic F of the pseudo auditory weight synthesis filters 114 and 115 by the unit 111
(Z), and these pseudo auditory weight synthesis filters 1
By performing filtering on the adaptive code vector and the pitch pulse train according to 14, 115, it becomes possible to search for a more suitable phase shift amount with less audible distortion than in the past.

Hereinafter, the present embodiment will be described in detail. First, an input audio signal from an audio input terminal is analyzed by a linear prediction analysis unit 102, and a linear prediction coefficient (LPC coefficient) is extracted. The extracted linear prediction coefficients are quantized by the synthesis filter information quantization unit 103, and the quantized linear prediction coefficients, that is, the code A indicating the synthesis filter information indicating the characteristics of the synthesis filter unit 104 is multiplexed by the multiplexing unit 507. Is output to The quantized linear prediction coefficients are used as filter coefficients.

The adaptive codebook 101 stores the past encoded excitation signal, and converts the past excitation signal into a pitch cycle L
, An adaptive code vector is generated.
When searching for the pitch period L, the distortion of the synthesized waveform obtained by synthesizing the adaptive code vector candidate corresponding to the pitch period candidate as speech by the synthesis filter 104 via the gain multiplication unit 108 and the addition unit 109 is output via the subtraction unit 122. The weighted distortion calculator 105 calculates the weighted distortion scale. The pitch period at which the distortion of the synthesized waveform with auditory weight becomes smaller is determined by the code selector 505.
Searched by. The pitch period L thus found
Is output to the multiplexing unit 107.

The noise codebook 120 stores a plurality of noise vectors, and outputs a noise vector corresponding to the noise code C searched by the code selection unit 106 to the phase adaptation unit 119. Phase adaptation section 119 cuts out a noise vector output from noise codebook 120 from a position corresponding to the phase shift amount provided from phase search section 112. The pitch periodizing section 118 gives the noise vector pitch periodicity by periodicizing the cut-out noise vector with the pitch period L, and outputs this as a noise code vector.

At the time of searching for the noise code C, a noise code vector candidate corresponding to the noise code candidate is
9, the distortion of the synthesized waveform obtained by synthesizing as speech by the synthesis filter 104 through the pitch periodizing unit 118, the multiplying unit 110, and the adding unit 109 is subjected to the auditory weighting by the auditory weighting distortion calculator 105 via the subtractor 122. Calculated on the distortion scale. Then, the code selecting unit 106 searches for a noise code C in which the distortion of the synthesized waveform with auditory weights becomes smaller. The noise code C searched in this way is output to multiplexing section 107.

The gain codebook 121 is the adaptive codebook 101
G0 applied to the adaptive code vector output from
And a gain G1 candidate to be applied to the noise code vector output from the pitch periodizing unit 118.

At the time of searching for a gain code, the addition unit 109 adds the adaptive code vector and the noise code vector obtained by multiplying the gain G0 and G1 candidates by the gain multiplication units 108 and 110, and adds the excitation code candidate. Is calculated by the auditory weighted distortion calculator 105 via the subtractor 122 using the auditory weighted distortion scale. Then, the code selection unit 106 searches for a gain code that reduces the distortion of the synthesized waveform with auditory weight. The gain code G searched in this way is output to multiplexing section 107.

The code of the pitch period L, the noise code C and the gain code G finally selected by the code selection unit 106 are multiplexed by the multiplexing unit 107 together with the code A representing the synthesis filter information. It is transmitted to the decoding side as a code stream. Excitation vectors generated using codes L, C, and G are stored in adaptive codebook 101 in preparation for encoding of excitation signals in the next section.

In the speech coding system based on the CELP method, as described above, a distortion measure with an auditory weight is used for selecting codes L, C, and G from which a sound source signal is generated, so that encoding distortion is reduced by an auditory sense. To make it harder to hear. The perceptual weight characteristic used in encoding is obtained from the linear prediction coefficient before quantization obtained by the linear prediction analysis unit 102. Since the information on the perceptually weighted distortion calculated by the perceptually weighted distortion calculator 105 is used only for code selection performed on the encoding side, there is no need to transmit this perceptual weighting characteristic information to the decoding side.

Next, a description will be given of a phase search method in the phase search unit 112 using the pseudo auditory weight synthesis filter which is a feature of the present invention. The phase search unit 112 synthesizes a pulse train (hereinafter referred to as a pitch pulse train), which is set at intervals of the pitch period L from the pulse train setting unit 113, with a synthesis filter 114 having a transfer characteristic represented by F (z), The adaptive code vector from adaptive codebook 101 is
A synthesis filter 115 having a transfer characteristic represented by (z)
By using the phase candidate setting unit 116 and the optimal phase selection unit 117 to search for the phase of the pitch pulse train that minimizes the error between the waveform and the waveform synthesized in step 2, the phase shift amount is obtained.

In this case, first, the pseudo-auditory weight synthesis filter setting unit 111 sets the synthesis filter information quantization unit 103.
The transfer characteristics of the synthesis filters 114 and 115 that perform pseudo auditory weighting are obtained by using the synthesis filter information that is the quantized linear prediction coefficients from. An example of this method will be described with reference to FIG.

FIG. 2A shows a synthesis filter 201 having a transfer characteristic of 1 / Aq (z), which is a normal 1 / Aq (z), using the quantized linear prediction coefficient αq (i) based on the prior art as a filter coefficient as it is. 4 (corresponding to the synthesis filters 514 and 515 in FIG. 4). On the other hand, FIG.
(B) is a pseudo-auditory weight synthesis filter 202 (the synthesis filter 1 in FIG. 1) having a transfer characteristic of F (z) according to the present invention.
14 and 115). Usually, the characteristic W (z) of the auditory weight filter used in the CELP method is represented by W (z) = A (z / γ1) / A (z / γ2) (1) Here, 0 <γ2 <γ1 <1, and the values of γ2 and γ1 are set according to the implementation of encoding. As an example, γ2 = 0.5 and γ1 = 0.98 can be used. A (z) is represented by A (z) = 1 + Σα (i) z ⁻ⁱ (2) using the unquantized linear prediction coefficient α (i).

On the other hand, the pseudo-auditory weight synthesis filter 202 (the synthesis filters 114 and 11 in FIG. 1) used in the present embodiment.
The characteristic Wq (z) of the pseudo auditory weight filter based on 5)
Is represented by the following equation. Wq (z) = Aq (z / γ1) / Aq (z / γ2) (3) where Aq (z) is a quantized linear prediction coefficient αq
Using (i), Aq (z) = 1 + Σαq (i) z- ⁱ (4)

The transfer characteristic F (z) of the pseudo-auditory weight synthesis filter 202 (the synthesis filters 114 and 115 in FIG. 1) used for the phase search in this embodiment is the characteristic Wq (z) of the pseudo-auditory weight filter described above. ) And the transfer characteristic 1 / Aq (z) of the ordinary synthesis filter are combined, and as shown in FIG. 2B, F (z) = [1 / Aq (z)] Wq (z) = [1 / Aq (z)] [Aq (z / γ1) / Aq (z / γ2)] (5)

Normally, due to the quantization error of the linear prediction coefficient, in other words, the quantization error of the synthesis filter information, the characteristic Wq (z) of the pseudo auditory weight filter and the weight characteristic W (z) of the auditory weight filter are the same. However, since encoding is performed on the encoding side so as to reduce the quantization error, 4
In low-rate encoding of the order of kbit / s, Wq
(Z) and the actual W (z) show very similar characteristics. This embodiment is characterized in that a phase search can be performed based on a distortion measure that reduces auditory distortion as compared with the related art using this fact.

The quantized linear prediction coefficient αq (i)
The transfer characteristic F (z) of the pseudo auditory weight synthesis filter using
Can be set, so that the same phase shift algorithm as the encoding side can be reproduced on the decoding side by using the same phase search algorithm as the encoding side.

According to the present invention, if the characteristic Wq (z) of the pseudo auditory weight filter is close to the characteristic of the auditory weight composite filter characteristic W (z) and can be reproduced based on the quantized synthetic filter information, The pseudo auditory weight filter is given by equation (3)
Needless to say, a material having a characteristic different from that described above is also effective. For example, even if Wq (z) is set to Wq (z) = Aq (z) / Aq (z / γ3) (6), the characteristics of the auditory weight filter can be reflected in a pseudo manner. A value in the range of about 0.7 to 0.9 is appropriate as the value of γ3. At this time, the characteristic F (z) of the pseudo-auditory weight synthesis filter used for the phase search is such that Fq (z) = [1 / Aq (z)] Wq (z) because the denominator and numerator Aq (z) are canceled. (7) = 1 / Aq (z / γ3) In this case, the pseudo auditory weight characteristic slightly changes as compared with the case of using F (z) in Expression (5), but is a sufficient characteristic for searching for a phase shift amount in which distortion is hardly perceptible. Further, since the order of F (z) is reduced, the filtering calculation itself is simplified, and there is an effect that the amount of calculation required for the phase search can be reduced.

Next, an example of a specific phase search method in this embodiment will be described. A pulse train (pitch pulse train) set at intervals of a pitch period L by the pulse train setting unit 113 is used as a pseudo-auditory weighting synthesis filter 1 having a transfer characteristic F (z).
A pseudo auditory weighting synthesis filter having a transfer characteristic F (z) is obtained by adding a phase candidate φ to the waveform synthesized by the phase candidate setting unit 116 and shifting the phase by shifting the adaptive code vector from the adaptive codebook 101. The error (waveform distortion) between the waveform synthesized in 115 and the optimal phase selecting unit 11
7 to obtain an auditory optimal phase shift amount φopt.

Here, 1 / Aq (z) is used as the characteristic of the filter used for the phase search in the prior art, whereas the pseudo-auditory weight synthesis filters 114 and 11 are used in the present invention.
5, the characteristic F (z) is used. For this reason, the conventional technique can be used as it is for the method of calculating the distortion by changing the phase φ with respect to the filtered waveform. As described below, by calculating the distortion when filtering is performed without actually performing filtering, an optimum phase can be searched.

Now, let x be an adaptive code vector, let L be a pitch period corresponding to the adaptive code vector x, and let yφ be a pulse train vector obtained by giving a phase φ to a pitch pulse train set at intervals of the pitch period L. The pulse train vector yφ is given by yφ = [yφ (0), yφ (1),..., Yφ (N−
1)] ^t can be defined. Let yφ (i) represent the i-th element of the vector yφ. Here, a vector y0 corresponding to φ = 0 is represented by y0 = [1,0,..., 0,1,0,.
...] ^t , and y0 (0) = 1. At this time, one element is arranged at intervals of the pitch period L.

On the other hand, since the pulse vector y1 corresponding to the phase φ = 1 is obtained by shifting the phase of the pulse vector y0 to the right by one sample, y1 (0) = 0 and y1 (1) = 1. , Y1 = [0,1,0, ..., 0,1,0, ...., 0,1,
0, ...] ^t . When the filtering operation of the pseudo auditory weight synthesis filters 114 and 115 is represented by a matrix F, the distortion Eφ at the pseudo auditory weight synthesis level between the adaptive code vector x and the pulse train vector yφ is: Eφ = (Fx−gφFyφ) ^t (Fx-gφFyφ) (8) Here, gφ is a gain value. E
Since the optimal gain Jifai when minimizing the φ is ^{gφ = (x t F t Fyφ} ) / (yφ t F t Fyφ) (9), when it is substituted into equation (6), the minimum value of E? ( E?) min is ^{(Eφ) min = x t F} t Fx - a ^{(x t F t Fyφ) 2} / (yφ t F t Fyφ) (10). Therefore, the second term on the right side of the equation (10) (x ^t F ^t F
^{yφ) 2 / (yφ t F} t Fyφ) to it can be seen that φ maximizing becomes optimum phase shift amount Faiopt.

As a method of reducing the calculation amount of the phase search, F
^t Fx calculated results corresponding to the vector r (= F ^t Fx)
Is obtained first, and then r ^t yφ is obtained for each phase candidate φ, whereby the calculation amount of the numerator part of the second term on the right side of Expression (10) can be reduced. The right side denominator of the second term of formula (10), and ^{dφ = Fyφ, Bφ = yφ t} F t Fyφ = d
Assuming that φ ^t dφ, the recursive formula Bφ = Bφ + 1 + d0 (N−1−φ) ² (11) holds for the calculation of the denominator term Bφ. By utilizing this relationship, the expression (10) can be calculated by one product-sum operation per one phase candidate.
Can be calculated in the denominator part of the second term on the right-hand side of, so the calculation amount is reduced to 1 / N.

The pseudo auditory weight synthesis filter 114,
As a method for further reducing the calculation amount of the filtering operation F of 115, the impulse response fn of the filters 114 and 115 is obtained up to a predetermined K samples, and by setting fn = 0 (n> K), the truncated impulse response is obtained. It is also effective to search for the amount of phase shift using this. Normally, the impulse response fn of the pseudo auditory weighting synthesis filters 114 and 115 has a large attenuation when n is about 10 samples, and the effect of the truncation on the characteristics becomes very small. For this reason, the pseudo auditory weight synthesis filters 114, 1
Fifteen impulse responses can be terminated when K = about 8 to 15. In this way, the pseudo-auditory weighting filter 114,
At 115, the calculation of filtering required for synthesis can be significantly reduced.

[Speech Decoding System] FIG.
FIG. 1 is a block diagram showing a configuration of a speech decoding system to which a speech decoding method according to the present embodiment is applied.
Has a configuration corresponding to the speech encoding system shown in FIG.

First, the code stream input from the code input terminal is input to the demultiplexing unit 307, where the code A, pitch period L code, noise code C, and gain code G of the quantized synthesis filter information are separated. Is done. Decoding is performed based on each of these codes.

The synthesis filter information decoding section 303 obtains a quantized linear prediction coefficient (LPC coefficient) based on the synthesis filter information given by the code A, and uses this as a filter coefficient of the synthesis filter section 304. .

The adaptive codebook 301 stores a past encoded excitation signal, and generates an adaptive code vector by repeating the past excitation signal at a pitch period L.
The phase search unit 312 is a phase search unit 11 on the encoding side in FIG.
As in the second embodiment, the adaptive codebook 301 includes a pulse train setting unit 313, pseudo auditory weighting synthesis filters 314 and 315, a phase candidate setting unit 316, and an optimal phase selecting unit 317.
The phase shift amount is obtained using the adaptive code vector and the pitch period L from the above, and is output to the phase adaptation section 319. That is, the pseudo-auditory weighting synthesis filter setting unit 311 uses the quantized synthesis filter information from the synthesis filter information decoding unit 303 to change the characteristics of the pseudo-auditory weighting synthesis filters 314 and 315 in the same procedure as on the encoding side. Then, the phase shift amount is obtained using this.

Noise codebook 320 outputs a noise vector corresponding to noise code C from demultiplexing section 307. Phase adaptation section 319 cuts out a noise vector output from noise codebook 320 from a position corresponding to the amount of phase shift given from phase search section 312. Pitch period 318
, The extracted noise vector is cycled by the pitch period L, and this is output as a noise code vector.

The gain codebook 321 is the adaptive codebook 301
G0 applied to the adaptive code vector output from
And a candidate for a gain G1 to be applied to the noise code vector output from the pitch periodizing unit 318, and output these gains G0 and G1 corresponding to the gain code G from the demultiplexing unit 307. .

The adaptive code vector and the noise code vector are multiplied by gains G0 and G1 in multipliers 308 and 310, respectively, and then added in adder 309 to form a sound source vector. The sound source vector is synthesized by the synthesis filter 304, and an audio signal is reproduced. The reproduced audio signal is appropriately passed through a post filter 330 and output as a final audio signal in order to further increase the subjective quality.

[0050]

As described above, according to the speech encoding / decoding method of the present invention, a pseudo-auditory weight synthesis filter for performing pseudo-auditory weighting from synthetic filter information is constructed. By searching for a phase shift amount for extracting a noise vector output from the noise codebook at a synthetic waveform level with a pseudo auditory weight using a filter, coding distortion is audibly reduced even at a low rate of about 4 kbit / s. It becomes possible to perform encoding / decoding of a high-quality excitation signal that is hardly perceived.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech encoding system according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram of a filter used in the present invention and a conventional phase search.

FIG. 3 is a block diagram showing the configuration of the speech decoding system according to the embodiment;

FIG. 4 is a block diagram showing a configuration of a conventional speech encoding system.

[Explanation of symbols]

Reference Signs List 101: adaptive codebook 102: linear prediction analysis unit 103: synthesis filter information quantization unit 104: synthesis filter unit (first synthesis filter) 105: perceptually weighted distortion calculation unit 106: code selection unit 107: multiplexing unit 108 110, a gain multiplying unit 109, an adding unit 111, a pseudo auditory weight synthesis filter setting unit 112, a phase search unit 113, a pulse train setting unit 114, 115, a pseudo auditory weight synthesis filter (second synthesis filter) 116, a phase candidate setting Unit 117: Optimal phase selection unit 118: Pitch periodicization unit 119: Phase adaptation unit 120: Noise codebook 121 ... Gain codebook 122 ... Subtraction unit 201 ... Synthesis filter 202 ... Pseudo-auditory weight synthesis filter 301 ... Adaptive codebook 303 ... Synthesis filter information decoding section 304... Synthesis filter section (first synthesis filter) 307 ... Demultiplexing sections 308 and 310... Gain multiplication section 309... Addition section 311... Pseudo-auditory weight synthesis filter setting section 312. 316: phase candidate setting unit 317: optimum phase selection unit 318: pitch period 319: phase adaptation unit 320: noise codebook 321: gain codebook 330: post filter

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-7-92999 (JP, A) JP-A-6-208399 (JP, A) JP-A-8-6597 (JP, A) JP-A-8-98 44400 (JP, A) NTT RD VOL. 45, NO. 4 pp. 325-330 1996 "Basic Algorithm of CS-ACELP" (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/12

Claims (4)

(57) [Claims] (A) generating an adaptive code vector by repeating, at a pitch cycle, a drive excitation signal output from an adaptive codebook storing past drive excitation signals; and (b) repeating at the pitch cycle. Phase shift amount for extracting a noise vector output from a noise codebook storing a plurality of noise vectors so that an error between a waveform synthesized from a pulse train to be synthesized and a waveform synthesized from the adaptive code vector is smaller. (C) generating a noise code vector by periodicizing the noise vector cut out in accordance with the phase shift amount with the pitch period; and (d) analyzing the input speech signal. The first synthesis filter whose characteristics are determined in accordance with the synthesis filter information obtained from the result is combined with the adaptive code vector and the noise code vector. Generating a synthesized speech signal by driving using a vector; (e) encoding synthesis filter information representing characteristics of the first synthesis filter; (f) auditory weighting of the synthesized speech signal Searching for the pitch period and the noise vector at which the subsequent distortion is smaller, wherein the phase shift amount searching step includes a second synthesizing filter for synthesizing speech by performing pseudo auditory weighting. An error between a synthesized waveform synthesized from the pulse train by the second synthesis filter and a synthesized waveform synthesized from the adaptive code vector by the second synthesis filter is reduced based on the synthesis filter information. A speech encoding method characterized by searching for the amount of phase shift as described above. 2. The speech encoding method according to claim 1, wherein an impulse response of said second synthesis filter is truncated at a finite number of predetermined samples. (A) generating an adaptive code vector by repeating a drive excitation signal output from an adaptive codebook storing past drive excitation signals at the pitch cycle; and (b) generating an adaptive code vector at the pitch cycle. A phase shift for cutting out a noise vector output from a noise codebook storing a plurality of noise vectors so that an error between a waveform synthesized from a repeated pulse train and a waveform synthesized from the adaptive code vector becomes smaller. (C) generating a noise code vector by periodicizing the noise vector cut out in accordance with the phase shift amount with the pitch period, (d) according to the synthesis filter information Driving the first synthesis filter whose characteristics have been determined using the adaptive code vector and the noise code vector. Generating a reproduced audio signal, wherein the phase shift amount searching step configures a second synthesis filter for synthesizing audio by performing pseudo auditory weighting from the synthesis filter information, 2. The phase shift amount is searched such that an error between a synthesized waveform synthesized from the pulse train by the second synthesis filter and a synthesized waveform synthesized from the adaptive code vector by the second synthesis filter becomes smaller. Audio decoding method. 4. The speech decoding method according to claim 3, wherein an impulse response of said second synthesis filter is truncated at a finite number of predetermined samples.