JP3332132B2 - Voice coding method and apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method of voice coding by which a high-quality voice signal can be reproduced with a small amount of operation even at a low coding bit rate below about 4kbps. SOLUTION: The voice coding method which supplies the drive signals generated by using the drive vectors obtained from a drive vector coding book to a weighted synthesizing filter 113, thereby generates synthesized voice signals, searches the drive vector that minimizes the distortion of synthesized voice signals and outputs coding parameters representing the information on the above drive vector and the filter factor of the synthesizing filter 113 has plural adaptive coding books 110, 120, a noise coding book 121 and a signal classifier 130. When the nature of input voice signals is classified as to be periodical, or periodical and stationary by the signal classifier 130, the drive vectors obtained from the adaptive coding books 110, 120 are used to generate drive signals. Otherwise, the drive vectors obtained from the noise coding book 121 are used to generate drive signals.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio encoding method and apparatus capable of reproducing high quality audio at a low bit rate.

[0002]

2. Description of the Related Art As an effective coding method for coding telephone band voice at a transmission rate of about 4 kbps, C.P.
ELP (Code Excited Linear Prediction) is known. This method is roughly classified into a process of obtaining a speech synthesis filter that models a vocal tract from input speech divided in units of frames, and a process of obtaining a drive vector serving as a drive signal corresponding to an input signal of the filter. Among these processes, the latter passes a plurality of drive vectors stored in the codebook one by one through a speech synthesis filter, calculates distortion of an obtained synthesized speech signal vector, that is, an error with respect to an input speech signal, and calculates the distortion. Is a process of searching for a drive vector that minimizes. This process is called a closed loop search,
This is a very effective method for reproducing good sound quality even at an encoding bit rate of about kbps.

However, the coding bit rate is 4 kbps.
If it is lower than this level, there is a problem that it is impossible to reproduce sound of sufficient quality in this method. Another problem of this method is that a large amount of calculation is required for the closed loop search. Hereinafter, these problems will be specifically described.

[0004] Regarding the CELP system, MRSchroeder
and BSAtal, “Code Excited Linear Prediction (CEL
P): High Quality Speech at Very Low Bit Rates ”,
Proc.ICASSP, pp.937-940,1985 and WSKleijin, DJ
Krasinski et al. “ImprovedSpeech Quality and Effic
ient Vector Quantization in SELP ”, Proc.ICASSP, p
p.155-158, 1988.

The outline of the CELP system will be described with reference to FIG. The input terminal 460 receives an audio signal in frame units. This input audio signal is input to the linear prediction analysis unit 45.
At the same time as the analysis is performed at 0, the filter coefficient of the weighted synthesis filter 430 is obtained, and at the same time, it is input to the auditory weighting unit 440 to obtain a weighted input audio signal. The zero-state response of the weighted synthesis filter 430 is subtracted from the weighted input audio signal, and a target vector 480 is generated.

Next, drive vectors are read out one by one from the adaptive codebook 411, multiplied by a gain in a gain circuit 421, and then input as a drive signal to a weighted synthesis filter 430, whereby a synthesized speech is obtained. A signal vector is generated. The evaluation unit 470 evaluates the distortion of the synthesized speech signal vector, that is, the difference from the target vector 480, and searches the adaptive codebook 411 for a drive vector so as to reduce the distortion. Vector. Next, a second drive vector is similarly searched from the noise codebook 412 in consideration of the effect of the first drive vector. Finally, the first and second drive vectors are multiplied by the optimum gains in gain circuits 421 and 422, respectively, to generate drive signals. The contents of the adaptive codebook 411 are updated by this drive signal, and the update is prepared for the input of the audio signal of the next frame.

As described above, in the CELP system, the weighted synthesis filter 430 corresponds to all the drive vectors stored in the adaptive codebook 411 and the noise codebook 412.
To perform a closed loop search for searching for a drive vector that minimizes the distortion. According to this CELP system, 8k
At an encoding bit rate of about bps, a relatively good quality audio signal can be reproduced, but at a low encoding bit rate of about 4 kbps or less, the adaptive codebook 4 can be reproduced.
There is a problem that the quality of the reproduced audio signal is degraded because the cycle of updating the contents of the H.11 or the noise codebook 412 becomes longer or the bit allocation becomes smaller.

This problem will be described in more detail. In the configuration of FIG. 7, the driving signal source for driving the weighted synthesis filter 430 is the first-stage adaptive codebooks 411 and 2.
It has a two-stage configuration consisting of a noise codebook 412 at the stage, and when the input speech signal is periodic, the adaptive codebook 41
1 expresses the periodic component of the drive signal, and the noise codebook 412 expresses the residual component of the drive signal that cannot be expressed by the adaptive codebook 411 alone. The book 412 is configured to share roles.

In this configuration, the encoding bit rate is 8 kb.
At about ps, it works well and a high-quality reproduced audio signal is obtained. However, at a low coding bit rate of about 4 kbps or less, the cycle of updating the contents of the adaptive codebook 411 and the noise codebook 412 (generally called a subframe) becomes longer, and the number of allocated bits decreases. Adaptive codebook 411 cannot express the periodic component of the drive signal with sufficient accuracy. As a result, a periodic component remains in the residual component of the drive signal that cannot be expressed by the adaptive codebook 411, and this cannot be expressed by the noise codebook 412, so that the quality of reproduced sound is degraded.

In addition, the CELP method has a problem that a huge amount of calculation is required to execute a closed loop search. To solve this computational problem, RCRose and TPBa
Invented by rnwell III. This method is SEV
(Self Excited Vocoder), and its operation and performance are described in RCRose and TPBarnwell III: “Qu
quality Comparison of Low Complexity 4800 bps Self Ex
cited and Code Excited Vocoders ”, Proc.ofICASSP'8
7, pp. 1637-1640, 1987. Hereinafter, this S
The EV system will be briefly described.

FIG. 8 is a block diagram of a drive signal search section in the SEV type speech coding apparatus. As can be seen from a comparison with FIG. 7, a major difference between the CELP system and the SEV system lies in the method of configuring the driving signal of the synthesis filter. CEL
In the P system, a drive signal is configured using an adaptive codebook and a noise codebook, whereas in the SEV system, only one or two adaptive codebooks (referred to as long-term predictors in the literature) 500 and 510 Is used to form a drive signal, and a special configuration of the adaptive codebook is used to reduce the amount of calculation. That is, each drive vector in adaptive codebooks 500 and 510 is configured to be shifted from the previous drive signal by a delay corresponding to the pitch period, and as a result, each drive vector has a configuration in which its elements overlap. Utilizing that
The calculation amount is reduced to about 1/20 by recursively calculating the output of the weighted synthesis filter 520 for the drive vector.

However, in the SEV system, the drive signal is always generated from the past history using the adaptive codebook regardless of the properties of the input speech signal. Generates an appropriate drive signal in areas where the characteristics of speech change, such as when there is little correlation, in the rising or falling part of the signal, in the transition from unvoiced to voiced, or in the transition from voiced to unvoiced. Therefore, there is a problem that the sound quality is further deteriorated as compared with the CELP system.

[0013]

As described above, among the conventional speech coding systems, the CELP system cannot reproduce high quality speech signals at a low coding bit rate of about 4 kbps or less. The SEV method has a problem that the amount of calculation is large, and the amount of calculation is reduced in the SEV method, but the quality of the reproduced audio signal is further deteriorated as compared with the SELP method.

An object of the present invention is to provide a speech encoding apparatus and apparatus capable of reproducing a high-quality speech signal with a small amount of computation even at a low encoding bit rate of about 4 kbps or less.

[0015]

In order to solve the above-mentioned problems, the present invention generates a drive signal using drive vectors obtained from a plurality of drive vector codebooks, and converts the drive signal into an analysis result of an input speech signal. The driving coefficient is supplied to a synthesis filter whose filter coefficient is determined based on the driving vector codebook. In a speech coding method for outputting coding parameters representing information of filter coefficients of a filter, at least two adaptive codebooks , at least one fixed codebook and at least one noise codebook are prepared as drive vector codebooks. , The first stage
1 of the adaptive codebook and the fixed codebook.
The codebook with the smaller signal distortion is selected.
Whether the first adaptive codebook was selected in one step or the fixed codebook
Determines whether or not it has been selected and cycles the properties of the input audio signal
Classification based on whether it is periodic or periodic and stationary
However, when the property of the input speech signal is periodic or periodic and stationary, a drive signal is generated using a drive vector obtained from at least two adaptive codebooks, and the property of the input speech signal is aperiodic or In a case where the driving signal is periodic and non-stationary, a driving signal is generated using at least a driving vector obtained from a noise codebook.

A speech coding apparatus according to the present invention comprises at least two adaptive codebooks prepared as a driving vector codebook , at least one fixed codebook and at least one noise codebook , Adaptive codebook and fixed
In the fixed codebook, the one in which the distortion of the synthesized speech signal is smaller
Codebook, and the first adaptive codebook is selected in the second stage.
Classifying means for judging whether or not the input speech signal is periodic or periodic and stationary based on whether the input speech signal is periodic or periodic and stationary. If the input speech signal is classified as periodic or periodic and stationary, a drive signal is generated using a drive vector obtained from at least two adaptive codebooks, and the property of the input speech signal is non-periodic or periodic and non-stationary. And a driving signal generating means for generating a driving signal using at least a driving vector obtained from the noise codebook.

As described above, in the present invention, the drive signal of the synthesis filter is formed by a linear combination of the drive vectors stored in the multi-stage drive vector codebook. Then, the characteristics of the input speech signal are classified according to whether they are periodic or whether they are periodic and stationary. If the characteristics are classified as periodic or periodic and stationary, the characteristics are obtained from at least two stages of adaptive codebooks. It generates a drive signal with a drive vector which is non-periodic
Or if it is classified as periodic and non-stationary,
Both are driven using the drive vector obtained from the noise codebook
Generate a signal .

As described above, in the conventional CELP system, when the input speech signal is periodic, the periodic component of the drive signal is expressed by the first-stage adaptive codebook, and can be expressed only by this adaptive codebook. Since the residual component of the drive signal that did not exist is represented by the second stage noise codebook, the cycle of updating the code of the adaptive codebook or the noise codebook is long at a low coding bit rate of about 4 kbps or less. And the number of allocated bits decreases. Therefore, the adaptive codebook cannot express a periodic signal with sufficient accuracy, and a periodic component remains in a residual component of the adaptive codebook,
Due to the inability of the noise codebook to represent the remaining periodic component, the quality of the reproduced audio signal has deteriorated.

On the other hand, in the present invention, when the input speech signal is periodic or periodic and stationary, the first adaptive codebook of the first stage expresses the periodic component of the drive signal. Even if a periodic component remains in the residual of the first stage due to a decrease in the encoding bit rate, this periodic component can be expressed by the second adaptive codebook of the second stage. Therefore, a high-quality reproduced audio signal can be obtained even at a low encoding bit rate.

On the other hand, when the input speech signal is non-periodic or periodic and non-stationary, the present invention uses at least a noise codebook and at most one adaptive codebook to generate the drive signal. Is not periodic, or a signal that is not stationary even if it is periodic and has a small correlation with the past drive signal, does not use the adaptive codebook that uses the correlation with the past drive signal, Since a configuration using a codebook can be used, a good reproduced audio signal can be obtained.

As described above, according to the present invention, the drive signal is composed of a plurality of stages of adaptive codebooks or a noise codebook depending on whether the input speech signal is periodic or periodic and stationary. By switching the configuration, a drive signal suitable for the properties of the input audio signal can be generated, so that a high-quality reproduced audio signal can be obtained even at a low encoding bit rate.

[0022]

[0023]

[0024]

BEST MODE FOR CARRYING OUT THE INVENTION

(First Embodiment) FIG. 1 is a block diagram showing a configuration of a speech coding apparatus according to a first embodiment of the present invention, in which a speech signal input terminal 100, a buffer 101, an LPC analysis unit 102, an LSP Quantizer 103, subframe division circuit 104, weighting filter 105, first adaptive codebook 110 constituting first-stage drive signal source, gain circuit 11
1, an adder 112, a weighted synthesis filter 113, a subtractor 114, a second adaptive codebook 120 and a noise codebook 121 constituting a second-stage drive signal source, a switch 122, a gain circuit 123, a gain codebook 124, signal classifier 13
0, search units 140, 141, 142 for searching the codebooks 110, 120, 121, 124, a multiplexer 150, and an output terminal 151 for encoding parameters.

Here, the signal classifier 130 classifies the property of the input speech signal according to whether it is periodic or periodic and stationary and outputs a classification determination signal. Further, the switcher 122 uses the second adaptive codebook 1 as a second-stage drive signal source based on the classification determination signal from the signal classifier 130.
20 and any one of the random codebook 121.

Next, the operation of the speech coding apparatus according to the present embodiment will be described. First, a digitized audio signal is input from an input terminal 100, is divided into intervals called frames (e.g., 20 ms), and is divided into buffers 10.
Stored in 1. The input audio signal read out from the buffer 101 in frame units is subjected to linear prediction analysis by the LPC analysis unit 102 in frame units, and a linear prediction coefficient a _i which is a parameter representing a spectrum envelope of the input audio signal.
(I = 1,..., P) are calculated and the LSP quantizer 103
Is input to

The LSP quantizer 103 converts the linear prediction coefficients into LSP (line spectrum pair) parameters,
Quantization is performed with a predetermined number of bits. The LSP parameters quantized by the LSP quantizer 103 are input to the multiplexer 150, and are also decoded and inversely transformed into linear prediction coefficients, and the quantized linear prediction coefficients aq
_i (i = 1,..., p) is obtained. Known methods can be used for the method of linear prediction analysis, the method of obtaining LSP parameters, and the method of quantizing LSP parameters. The quantized linear prediction coefficients are assigned to a weighting filter 1
05 and the weighted synthesis filter 113.

On the other hand, the input audio signal read out from the buffer 101 on a frame basis is
Each frame is divided into a plurality of sections called subframes by 04, and is input to the weighting synthesis filter 105 in subframe units to become a target vector for driving signal search. Then, the weighted synthesis filter 11
The search processing of the drive signal for driving No. 3 is performed on a subframe basis as follows.

In the speech coding apparatus of the present embodiment, the first adaptive codebook 110 is used as a drive signal source in the first stage, and one of the second adaptive codebook 120 and the noise codebook 121 is used in the second stage. It has a two-stage configuration as a drive signal source, and a drive vector is sequentially searched from the codebook of each stage. This search procedure will be described with reference to the flowchart shown in FIG.

First, a drive vector search is performed for the first adaptive codebook 110 in the first stage. That is, when an audio signal is input from the input terminal 100 (step S1),
Prior to searching adaptive codebook 110, weighting filter 1
At 05, the input speech signal is subjected to perceptual weighting, the target vector X is generated by subtracting the influence of the weighting synthesis filter 105 from the previous subframe (step S2), and the first adaptive codebook 110 is searched (step S2). Step S3). The first adaptive codebook 110 is used to represent a periodic component (pitch) of a speech signal, and a driving vector e (n) stored in the adaptive codebook 110 is expressed by the following equation. It is created by cutting out the past drive signal by the subframe length.

E (n) = e (nL), n = 1,..., N (1) where L is a lag and N is a subframe length. The search for the first adaptive codebook 110 is performed by the search unit 140, which is a distortion of the synthesized speech signal vector obtained by passing the drive vector e through the weighted synthesis filter 113, that is, an error between the target vector X and the synthesized speech signal vector. The search is performed by searching for a lag that minimizes according to a known method. The lag can be in whole or fractional samples.

The weighting filter 105 and the weighting synthesis filter 113 can be formed by a known method. For example, if W (z) and Hw (z) are transfer functions, respectively, W (z) = A ( z / r) / A (z / r ') (2) Hw (z) = W (z) / Aq (z) (3) A (z) = 1 + Σa i z -i (4) Aq (z) = can be expressed as ^{_{1 + Σaq i z -i (5}} ).

Next, the search unit 14 searches the codebook in the second stage.
Perform by 1. In this case, first, the input speech signal to be encoded is classified by the signal classifier 130 according to whether it is periodic or whether it is periodic and stationary (step S4).
More specifically, classification is performed by a known voiced / unvoiced determination method, a method combining this with a determination of the continuity of the pitch period, or an autocorrelation method.

FIG. 2 is a diagram showing an example of the waveform of a voice signal in a voiced section and an unvoiced section. As shown in the figure, in a transitional part transitioning from an unvoiced section to a voiced section, and in a transitional part transitioning from a voiced section to a unvoiced section, a non-stationary transition in which the pitch period and amplitude change are periodic but not uniform It is in a state, and is in a periodic and steady state except for the transient part. In the simplest case, the signal classifier 130 may perform the classification based on whether the input voice signal is a voiced section or an unvoiced section based on whether it is periodic or not. However, the signal classifier 130 determines whether the voiced section is more stationary. Accordingly, the classification may be performed based on whether the information is periodic and stationary.

If the result of the classification in step S4 is that the input speech signal is periodic or periodic and stationary, the second adaptive codebook 120 is searched for as the second-stage codebook (step S5). Otherwise, the random codebook 121 is searched (step S6). The search for the second adaptive codebook 120 is performed in the same manner as the search for the first codebook 110. Also,
The search for the random codebook 121 is performed in the same manner as in the conventional CELP method. At this time, a known technique such as an overlapping codebook, backward filtering, and preliminary selection can be used to reduce the amount of calculation.

Finally, the gain codebook 124 is searched (step S7). The gain codebook 124 has, as a representative vector, a vector having a gain multiplied by a drive vector searched from the first and second codebooks as an element. Then, the driving vector searched from the first adaptive codebook 110 in the first stage and the first adaptive codebook 120 or the noise codebook 12 in the second stage extracted through the switch 122.
The gain vector 111,
The drive vector obtained by multiplying each of the gains 123 by the gain searched from the gain codebook 124 and further adding them by the adder 112 is added to the weighted combined vector 1
13 so that the distortion (error with respect to the target vector) of the synthesized speech signal vector obtained through step 13 is minimized.
At 42, a gain codebook 124 in a known manner in a closed loop
Is searched.

As described above, the search units 140, 141,
The index (first index) of the first adaptive codebook 110 searched in 142, the second adaptive codebook 120
Alternatively, the index (second index) of the noise codebook 121 and the index (third index) of the gain codebook 124 are the quantized parameters from the LSP quantizer 103 and the classification judgment from the signal classifier 130. Along with the signal, it is output via the multiplexer 150 as an encoding parameter. Multiplexer 15
The encoding parameter output from 0 is transmitted to a transmission path or a storage medium.

As described above, in the speech coding apparatus according to the present embodiment, if the input speech signal is classified as periodic or periodic and stationary by the signal classifier 130, the first
, The periodic component of the drive signal of the weighted synthesis filter 113 is represented by the adaptive codebook 110, and even if the periodic component remains in the residual obtained from the subtractor 114 due to the decrease in the encoding bit rate, Since this periodic component can be expressed by the second adaptive codebook 120, a high-quality reproduced audio signal can be obtained in the audio decoding device even at a low encoding bit rate.

If the input speech signal is classified as non-periodic or periodic and non-stationary by the signal classifier 130, the noise codebook 121 or the adaptive codebook 110
And a noise codebook 121 to generate a drive signal, it is possible to obtain a good reproduced audio signal in the audio decoding device without extensive use of an adaptive codebook utilizing correlation with past drive signals. .

Next, the speech decoding apparatus according to this embodiment will be described. FIG. 4 is a block diagram illustrating a configuration of the speech decoding device according to the present embodiment. This speech decoding apparatus includes an encoded data input terminal 160, a demultiplexer 161, an LSP inverse quantizer 162, a first adaptive codebook 1
70, gain circuit 171, adder 172, synthesis filter 173, post filter 174, second adaptive codebook 18
0, a noise codebook 181, a switch 182, a gain circuit 183, a gain codebook 184, and a reproduction audio signal output terminal 190.

The input terminal 160 receives coding parameters output from the speech coding apparatus shown in FIG. 1 via a transmission line or a storage medium. This coding parameter is input to the demultiplexer 161 and is quantized by the LSP quantizer 103 in FIG. 1, the LSP parameter, the index (first index) of the first adaptive codebook 110, and the second adaptive codebook. 120 or noise codebook 121
Index (second index), gain codebook 1
The 24 indices (third index) and the classification judgment signal from the signal classifier 130 are separated and decoded.

Among the outputs of the demultiplexer 161, the quantized LSP parameter is the LSP inverse quantizer 162
, The LSP parameters are restored, and the first index is assigned to the first adaptive codebook 170 and the second index is assigned to the second adaptive codebook 180 and the random codebook 1.
81, the third index is input to the gain codebook 184, and the classification determination signal is input to the switch 182.

The first adaptive codebook 170, the second adaptive codebook 180, the noise codebook 181 and the gain codebook 184
Has the same configuration as the first adaptive codebook 110, the second adaptive codebook 120, the noise codebook 121, and the gain codebook 124 in the speech coding apparatus shown in FIG.

The operation of the speech decoding apparatus will be described.
A first driving vector obtained by searching the first adaptive codebook 170 based on the first index output from the demultiplexer 161 is based on the third index from the gain codebook 184 in the gain circuit 171. After being multiplied by the searched gain, it is input to the adder 172. On the other hand, a second drive vector obtained by searching the second adaptive codebook 180 or the noise codebook 181 based on the second index is selected by the switch 182 based on the classification determination signal, and similarly, After being multiplied by the gain searched for based on the third index from the gain codebook 184 in the gain circuit 172, it is input to the adder 172.

The drive vector synthesized by the adder 172 is provided as a drive signal to a synthesis filter 173 whose transfer characteristic is controlled by the LSP parameter from the LSP dequantizer 162, and the synthesized filter 173 synthesizes a speech signal vector. Is generated. This synthesized speech signal vector is input to a post-filter 174, whose transfer characteristics are also controlled by LSP parameters, where known processing for improving the subjective quality of the reproduced speech, such as pitch emphasis, formant emphasis, and high-frequency emphasis After being subjected to processing such as gain adjustment and the like, it is output from the output terminal 190 as a reproduced audio signal. This reproduced audio signal has high quality as described above.

(Second Embodiment) FIG. 5 is a block diagram of a speech coding apparatus according to a second embodiment. The parts corresponding to those in FIG. 1 are denoted by the same reference numerals and the description will focus on the differences from the first embodiment. The difference between the second embodiment and the first embodiment is 2 In the method of selecting a codebook to be a drive signal source of the second stage. That is, in the first embodiment, the characteristics are classified by analyzing the input audio signal,
The codebook used as the second-stage drive signal is switched between the second adaptive codebook 120 and the noise codebook 121.

On the other hand, in the second embodiment, the first adaptive codebook 21 is used as the codebook serving as the first-stage drive signal source.
0, and a fixed codebook 211 in which the stored driving vector is always fixed.
One of 0 and 211 is selected based on the minimum distortion standard of the synthesized voice signal level, and from the result, the property of the input voice signal is similarly classified according to whether it is periodic or not, and whether it is periodic and stationary. The second adaptive codebook 120 and the random codebook 121 are used as codebooks used as the second-stage drive signal source.
Select one of In other words, in the first embodiment, the codebook is selected in an open loop, whereas in the second embodiment, the codebook in the second stage is selected in a closed loop.

The operation of this embodiment will be specifically described.
First, LPC analysis and LSP parameter quantization are performed in the same manner as in the first embodiment for each frame. Next, in search section 140, search for first adaptive codebook 210 is performed in the same manner as in the first embodiment, and search for fixed codebook 211 is performed in the same manner as in search for noise codebook 121 in the first embodiment. , And extracts the drive vector searched from the adaptive codebook 210 and the fixed codebook 211 via the switch 212.

The result of the search of the first adaptive codebook 210 and the fixed codebook 211, that is, the driving vector searched from the first adaptive codebook 210 is added to the gain circuit 1
11 and the drive vector obtained by multiplying the gain searched from the gain codebook 124 by the weighting synthesis filter 1
13 and a driving vector obtained by multiplying the distortion of the synthesized speech vector obtained through the loop 13 and the driving vector searched from the fixed codebook 211 by the gain searched from the gain codebook 124 by the gain circuit 111. The distortion of the synthesized speech vector obtained through step 113 is compared, and the drive vector that gives the minimum distortion is the drive vector from the first-stage drive signal source. The index of the first adaptive codebook 210 or fixed codebook 211 obtained by the search unit 140 by this search is
0 is provided to the signal classifier 230.

The signal classifier 230 determines whether the first adaptive codebook 210 has been selected as the first-stage drive signal source from the index input from the search unit 140 or not.
By determining whether 1 has been selected, the characteristics of the input audio signal are classified according to whether they are periodic or whether they are periodic and stationary. Then, the signal classifier 230
When the adaptive codebook 210 is selected, that is, when the input speech signal is periodic or periodic and stationary, the second adaptive codebook 220 is selected as the second-stage drive signal source, Otherwise, switch 122 is controlled so that random codebook 121 is selected.

The search for the codebook as the second-stage drive signal source is performed in the same manner as in the first embodiment. In this case, the signal classifier 230 selects the second-stage codebook in consideration of not only whether or not the first adaptive codebook 210 has been selected as the first-stage drive signal source but also the continuity of the pitch. May be performed.

The speech decoding device corresponding to the speech encoding device in FIG. 5 is the same as the speech decoding device in FIG. (Third Embodiment) In the first and second embodiments, the drive signal source is constituted by a two-stage codebook. However, the drive signal source may be constituted by three or more stages. FIG. 6 is a block diagram showing a configuration of a speech encoding device according to the third embodiment when the number of drive signal sources is three. As shown in the figure, in the present embodiment, another noise codebook 330 which is a third-stage drive signal source and a gain codebook 124 are added to the output of the configuration of the first embodiment shown in FIG. A gain circuit 331 for multiplying by the gain searched from the above and a search unit 143 for searching the noise codebook 330 are added.

The search operation of the codebook in this embodiment is as follows.
First, a first adaptive codebook 110 which is a drive signal source of the first stage
Is searched by the search unit 140, then the search for the second adaptive codebook 120 or the noise codebook 121, which is the second-stage drive signal source, is performed by the search unit 141, and finally the third-stage drive signal source The search is performed by the search unit 143 to search for the random codebook 330.

In the above embodiment, as means for classifying the properties of the input speech signal based on whether they are periodic or whether they are periodic and stationary, as shown in FIG. Classification is performed by voiced / unvoiced determination or the like, or whether the first adaptive codebook 210 is selected in the first step as shown in FIG. 5, or whether the second adaptive codebook 210 is selected in the second step as shown in FIG. Although the method of indirectly classifying based on whether or not the adaptive codebook 120 is selected is used, the present invention is not limited to this.

For example, the power of the input voice signal may be determined to classify the characteristics of the input voice signal according to whether it is periodic or periodic and stationary, or the first adaptive codebook 1 may be used.
Gain vector 111 to the drive vector read from
Various modifications are possible, such as determining whether the gain multiplied by exceeds a predetermined threshold, and classifying the property of the input audio signal as periodic or periodic and stationary when the threshold is exceeded. is there.

[0056]

As described above, according to the present invention, the characteristics of an input voice signal are classified according to whether they are periodic or whether they are periodic and stationary, and are classified as periodic or periodic and stationary. In this case, by generating a drive signal using the first and second two-stage adaptive codebooks, the periodic components that cannot be sufficiently removed by the first adaptive codebook and remain are removed by the second adaptive codebook. When classified as non-periodic or periodic and non-stationary, a driving signal is created using a noise codebook to represent a signal having a small correlation. As a result, a drive signal suitable for the properties of the input audio signal can always be formed, and high-quality audio can be reproduced even at a bit rate as low as about 4 kbps.

Further, in the present invention, since the drive signal is composed of an adaptive codebook, there is also an effect that the amount of calculation which has been a problem in the CELP system can be reduced by using a special configuration of the adaptive codebook. That is, similarly to the SEV method, each driving vector in the adaptive codebook is configured to be shifted from the past driving signal by a delay corresponding to the pitch period, and as a result, each driving vector has a configuration in which its elements overlap. By utilizing the fact that the calculation of the output of the synthesis filter with respect to the drive vector is performed recursively, it is possible to greatly reduce the amount of calculation.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech encoding device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of waveforms in a voiced section and an unvoiced section of an audio signal.

FIG. 3 is a flowchart showing a flow of main processing in the embodiment.

FIG. 4 is a block diagram showing the configuration of the speech decoding apparatus according to the embodiment;

FIG. 5 is a block diagram showing a configuration of a speech encoding device according to a second embodiment of the present invention.

FIG. 6 is a block diagram illustrating a configuration of a speech coding apparatus according to a third embodiment of the present invention.

FIG. 7 is a block diagram showing a typical configuration of a conventional CELP-based speech coding apparatus;

FIG. 8 is a block diagram showing a configuration of a main part of a conventional speech coding apparatus using the SEV method.

[Explanation of symbols]

 Reference Signs List 100 audio signal input terminal 101 buffer 104 subframe division circuit 102 LPC analysis section 103 LSP quantizer 105 weighting filter 110 first adaptive codebook 111 gain circuit 112 adder 113 weighted Synthesis filter 114 Subtractor 120 Second codebook 121 Noise codebook 122 Switcher 123 Gain circuit 124 Gain codebook 130 Signal classifier 140-143 Search unit 150 Multiplexer 151 Coding parameter Output terminal 210 first adaptive codebook 211 fixed codebook 150 multiplexer 160 input parameter input terminal 161 demultiplexer 162 LSP inverse quantizer 170 first adaptive codebook 171 gain circuit 173 Synthetic filter 174: Post filter 80 second adaptive codebook 181 noise codebook 182 switch 183 gain circuit 210 first adaptive codebook 211 fixed codebook 212 switch 230 signal classifier 330 noise codebook 331 Gain circuit

Claims (2)

(57) [Claims] A drive signal is generated using a drive vector obtained from a plurality of drive vector codebooks, and the drive signal is supplied to a synthesis filter whose filter coefficient is determined based on an analysis result of an input speech signal; The drive vector codebook searches for a drive vector that reduces the distortion of the synthesized speech signal output from the synthesis filter from the drive vector codebook, and outputs at least an encoding parameter representing information on the drive vector and the filter coefficient of the synthesis filter. In the speech coding method, at least two adaptive codebooks, at least one fixed codebook and at least one noise codebook are prepared as the drive vector codebook, and the first adaptive codebook and the fixed codebook are provided in a first stage. Of the codebooks, the one with the smaller distortion of the synthesized voice signal is selected,
In the second step, it is determined whether the first adaptive codebook or the fixed codebook has been selected in the first step, and whether the property of the input speech signal is periodic or not, and And if the property of the input speech signal is periodic or periodic and stationary, generate the drive signal using a drive vector obtained from at least the two adaptive codebooks; A speech coding method comprising: generating a drive signal using at least a drive vector obtained from the noise codebook when the property of the speech signal is non-periodic or periodic and non-stationary. 2. A drive signal is generated using drive vectors obtained from a plurality of drive vector codebooks, and the drive signal is supplied to a synthesis filter whose filter coefficient is determined based on an analysis result of an input speech signal. The drive vector codebook searches for a drive vector that reduces the distortion of the synthesized speech signal output from the synthesis filter from the drive vector codebook, and outputs at least an encoding parameter representing information on the drive vector and the filter coefficient of the synthesis filter. In the speech encoding device, at least 2 provided as the drive vector codebook
Two adaptive codebooks, at least one fixed codebook and at least one noise codebook, and the first adaptive codebook and the fixed codebook in the first stage, which codebook has a smaller distortion of the synthesized speech signal. Selected,
Classification means for judging whether a first adaptive codebook or a fixed codebook is selected in a second step, and classifying the input speech signal according to whether the property of the input speech signal is periodic or periodic and stationary. And if the property of the input audio signal is classified as periodic or periodic and stationary by the classification means,
At least the noise code book when the two using a driving vector obtained from the adaptive codebook generates the drive signal, the nature of the input speech signal is classified as aperiodic or periodic and unsteady And a drive signal generating means for generating the drive signal using a drive vector obtained from the audio encoding apparatus.