JPH05249999A - Learning type voice coding device - Google Patents

Abstract

PURPOSE:To provide a learning type voice coding device capable of composing a voice of higher quality at a limited bit rate as less than 8 kbps. CONSTITUTION:A learning type voice composing device has an adaptive code book 110 in which drive signal vectors are stored; a minimum distortion searching circuit 115 for searching an optimum drive signal vector from the adaptive code book 110 in reference to an input voice signal; a composing filter 112 for composing a voice signal by use of the searched optimum drive signal vector; a buffer 131 for accumulating the information of the searched optimum drive signal vector; a training vector forming part 132 for cutting the accumulated information of drive signal vector in a determined length to form a training vector; and a learning part 133 for successively correcting the drive signal vectors in the code book by use of the training vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder,
In particular, the present invention relates to a learning type speech coding apparatus suitable for coding a speech signal at a low bit rate of about 8 kbps or less.

[0002]

2. Description of the Related Art A technique for encoding a voice signal at a low bit rate with high efficiency is an important technique for effective use of radio waves and reduction of communication cost in mobile communications such as car telephones and in-house communications. is there. CELP (Code Ex) is used as a high-quality speech coding method at a bit rate of 8 kbps or less.
The cited Linear Prediction) method is known.

This CELP method is based on M.A. of AT & T Bell Labs.
“Code-Excited Lin” by R. Schroeder and BSAtal
ear Prediction (CELP) “High-Quality Speech at Very
Since it was announced in LowBit Rates "Proc.ICASSP; 1985, pp.937-939 (Reference 1), it has been attracting attention as a method of synthesizing voices of product quality, and various studies such as quality improvement and reduction of calculation amount have been made. The features of CELP method are LP
The drive signal of the C (Liner Predictive Coding) synthesis filter is stored in the codebook as the drive signal vector, and the optimum drive signal vector is searched from the codebook while evaluating the error between the synthesized voice signal and the input voice signal. There is a point to do.

FIG. 9 is a block diagram of a speech coder according to the latest CELP method. In the figure, the sampled audio signal sequence which is the input signal is the input terminal 60.
It is input from 0 in frame units. A frame consists of L signal samples, and when the sampling frequency is 8 kHz, L = 160 is generally used. Although not shown in FIG. 9, LPC analysis is performed on the input L-sample speech signal sequence prior to the search for the drive signal vector, and LPC prediction parameters {α ₁ , i = 1, 2, ...
p} is extracted. This LPC prediction parameter α ₁ is
It is supplied to the LPC synthesis filter 630. Note that p is the predicted order, and p = 10 is generally used. The transfer function H (z) of the LPC synthesis filter 630 is given by [Equation 1].

[0005]

[Equation 1]

Next, a process of searching for an optimum drive signal vector while synthesizing a voice signal will be described. First,
From the 1-frame audio signal input to the input terminal 600, the subtractor 610 performs the synthesis filter 630 for the previous frame.
The effect of the internal state of the current frame on the current frame is subtracted. The signal sequence obtained from the subtractor 610 is divided into four subframes, and becomes the target signal vector of each subframe.

The driving signal vector, which is the input signal of the LPC synthesis filter 630, is selected from the white noise codebook 710 and the driving signal vector selected from the adaptive codebook 640 multiplied by a predetermined gain in the multiplier 650. The obtained noise vector is multiplied by a predetermined gain in the multiplier 720 and is added by the adder 660.

Here, the adaptive codebook 640 is referred to as reference 1
The pitch prediction analysis described in the above is performed by closed loop operation or analysis by synthesis.For details, see WBKleijin DJ Krasinski and R.
H. Ketchum, "Improved Speech Quality and Efficient Ve
ctor Quantization in CELP ", Proc.ICASSP, 1988, pp.155
-158 (reference 2). According to this reference 2, the drive signal of the LPC synthesis filter 630 is set to the pitch search range a to b (a and b are sample numbers of the drive signal,
Normally, a delay circuit 670 is provided over a = 20, b = 147).
By delaying one sample at a time, a drive signal vector for the pitch period of a to b samples is created, and this is stored in the adaptive codebook as a codeword.

When the optimum drive signal vector is searched for, one codeword of the drive signal vector corresponding to each pitch period is read from the adaptive codebook 640 and is multiplied by a predetermined gain in the multiplier 650. . Then, filter calculation is performed by the LPC synthesis filter 630 to generate a synthesized voice signal vector. The generated synthesized speech signal vector is subtracted from the target signal vector by the subtractor 620. The output of the subtractor 620 is input to the miscalculation calculation circuit 690 via the perceptual weighting filter 680, and the mean square error is obtained. The information on the mean square error is further input to the minimum distortion search circuit 700, and its minimum value is detected.

The above process is performed for all the drive signal vector codewords in the adaptive codebook 640, and the codeword number giving the minimum value of the mean square error is obtained in the minimum distortion search circuit 700. The gain multiplied by the multiplier 650 is also determined so that the mean square error is minimized.

Next, the optimum white noise vector is searched for in the same manner. That is, the codewords of the noise vector are read one by one from the white noise codebook 710, multiplied by the gain in the multiplier 720, and subjected to the filter operation in the LPC synthesis filter 630 to generate a synthetic speech signal vector and to obtain the target. The calculation of the mean square error with the vector is performed for all noise vectors. Then, the number and the gain of the noise vector that gives the minimum value of the mean square error are obtained. The perceptual weighting filter 680 is used to shape the spectrum of the error signal output from the subtractor 620 and reduce the distortion known to humans.

As described above, in the CELP system, since an optimum drive signal vector that minimizes the error between the synthesized voice signal and the input voice signal is obtained, a high quality voice is synthesized even at a low bit rate of about 8 kbps. can do. However, it has been confirmed that at a bit rate of 8 kbps or less, the deterioration of quality is perceived because the number of bits allocated for encoding the drive signal becomes insufficient.

[0013]

As described above, the conventional CELP system can synthesize high-quality speech at a bit rate of about 8 kbps or higher, but at a bit rate lower than this, it is not suitable for driving signal coding. There is a problem in that the number of allocated bits is insufficient and the deterioration of quality is perceived, which is not practically sufficient.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a learning-type speech coder capable of synthesizing higher-quality speech at a limited bit rate of about 8 kbps or less. And

[0015]

SUMMARY OF THE INVENTION In order to solve the above problems, the present invention refers to a codebook (adaptive codebook) in which a drive signal vector is stored as a codeword, and an adaptive codebook with reference to an input voice signal. Search means for searching for an optimum driving signal vector, a synthesis filter for synthesizing a voice signal using the optimum driving signal vector searched by this searching means, and a training vector using the optimum driving signal vector It is characterized by comprising a training vector creating means and a learning means for sequentially correcting the drive signal vector in the codebook using the training vector created by this means.

[0016]

In the present invention, the optimum drive signal vector searched from the adaptive codebook, that is, the drive signal vector actually used for encoding by driving the synthesis filter is used, and this is used as a training vector in the adaptive codebook. Of the drive signal vector, specifically, the representative vector selected based on a predetermined standard among the drive signal vectors is sequentially corrected. This processing is performed in parallel with the encoding every time a new drive signal vector is searched.

By the learning process in which the drive signal vector is sequentially corrected in this way, the drive signal vector in the adaptive codebook is sequentially changed to a vector that can more accurately synthesize the voice of the speaker. As a result, high quality speech synthesis is possible even at a low bit rate of, for example, about 8 kbps or less.

[0018]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a learning-type speech encoding apparatus according to an embodiment of the present invention.

In FIG. 1, an audio signal sampled at a predetermined sampling frequency (for example, 8 kHz) is input to the input terminal 100 in units of frames. This input audio signal is first input to the frame buffer 101. In the frame buffer 101, the input audio signal sequence is cut out in units of L (for example, L = 160) samples, and 1
It is stored as a frame signal. Frame buffer 10
The 1-frame input audio signal from 1 is supplied to the LPC analysis circuit 102 and the weighting filter 106.

The LPC analysis circuit 102 uses an autocorrelation method, for example, for an LPC (Linear Predi
ctive Coding: Linear predictive coding) analysis is performed, and P L
The PC prediction coefficient {α ₁ , i = 1, 2, ... P,} or the reflection coefficient {k ₁ , i = 1, 2, ..., P} is extracted. The extracted prediction coefficient or reflection coefficient is encoded by the encoding circuit 103 with a predetermined number of bits, and then the weighting filter 106 and the weighting synthesis filters 107 and 11 are used.
2,122 is used.

The weighting filter 106 weights the input speech signal sequence when searching the drive signal vector of the synthesis filter from the adaptive codebook 110 and the noise codebook 120. The transfer function H (z) of the synthesis filter in the weighting synthesis filters 107, 112, 122 is described by [Equation 1]. At this time, the transfer function W (z) of the weighting filter 106 is represented by [Equation 2].

[0022]

[Equation 2] However, γ is a parameter for controlling the strength of weighting (0 ≦ γ ≦ 1).

Weighting synthesis filters 107, 112, 1
22 is a filter in which the synthesis filter of the transfer function H (z) shown in [Equation 1] and the weighting filter of the transfer function W (z) are connected in series, and the transfer function H _w (z) is [ [Equation 3].

[0024]

[Equation 3]

As in the present embodiment, the weighting filter 106
Is used, it is possible to reduce audible coding distortion. Further, in the present embodiment, the weighting filter 106 is provided outside the drive signal vector search loop, and as a result, the amount of calculation required for the search is substantially eliminated.

Furthermore, the weighting synthesis filters 112, 1
A weighting synthesis filter 107 having an initial memory is provided so that 22 does not affect the search of the drive signal vector. The weighting synthesis filter 107 is
Weighted synthesis filters 112, 12 at the end of the previous frame
2 has an internal state held by 2 as an initial state.

Then, a zero input response vector of the weighting synthesis filter 107 is created, and the subtracter 108 subtracts the zero input response vector from the output of the weighting filter 106. As a result, the weighting synthesis filter 11
The initial state of 2,122 can be set to zero, and the drive signal vector can be searched without considering the influence of the previous frame. The above processing is all performed in frame units. Next, a drive signal vector search process performed by dividing a frame into M (usually M = 4) subframes and performing the subframe unit will be described.

The search for the optimum drive signal vector is performed in the order of the adaptive codebook 110 and the noise codebook 120. The adaptive codebook 110 contains K dimensions (K = L /
128 drive signal vectors of M) are stored so as to correspond to a pitch period of 20 to 147 samples. In the search for the drive signal vector, first, the drive signal vector X _j specified by the index j described later is sequentially read from the adaptive codebook 110, the multiplier 111 multiplies X _j by a predetermined gain β, and then weighted synthesis is performed. It is supplied to the filter 112. Weighting synthesis filter 1
In step 12, the driving signal vector multiplied by the gain β is filtered to create a synthetic speech vector.

On the other hand, the input audio signal read from the frame buffer 101 is weighted by the weighting filter 106, and then the influence of the previous frame is subtracted by the subtractor 108. Using the audio signal vector Y output from the subtractor 108 as a target vector, the subtractor 11
In 3, the error vector E _j from the synthesized speech vector from the weighting synthesis filter 112 is calculated. Then, the squared error calculation circuit 114 calculates the sum of squared error ‖E _j ‖,
The minimum value of this ‖E _j ‖ and the index j that gives the minimum value are detected by the minimum distortion search circuit 115. This index j is the adaptive codebook 110 and the multiplexer 1
42.

Specifically, the error vector E _j is represented by, for example, [Equation 4]. This error vector ‖E _j ‖ is β
By partially differentiating and setting it as zero, the minimum value of ‖E _j ‖ when β is optimized is expressed by [Equation 5]. However, β
Is a gain given by the multiplier 111.

[0031]

[Equation 4]

[0032]

[Equation 5]

Where ‖X‖ is the square norm, (X, Y)
Represents an inner product, and H is an impulse response matrix of a weighting synthesis filter (transfer function: H _w (z)) given by [Equation 6].

[0034]

[Equation 6]

As is clear from [Equation 5], the search for the drive signal vector from the adaptive codebook 110 calculates the second term on the right-hand side of [Equation 5] for all codewords X _j , which is the maximum. This is done by detecting the index j that becomes

When the optimum drive signal vector X _opt is searched from the adaptive codebook 110 in this way, the subtracter 113 subtracts the output of the weighting synthesis filter 112 corresponding to X _opt from the target vector Y, and this subtraction is performed. The output of the unit 113 is used as the target vector for the noise vector search from the noise codebook 120. Noise Codebook 1
The search for noise vectors from 20 is also an adaptive codebook 1
The search for the drive signal vector from 10 can be performed in exactly the same way. If the code vector obtained by the search from the noise vector 120 is N _opt , the drive signal vector X of the synthesis filter is

[0037]

[Equation 7] Is expressed as However, β and g are subtractors 111 and 1 respectively.
21 is a gain given to the drive signal vector and the noise vector searched from the adaptive codebook 110 and the noise codebook 120 in FIG.

The drive signal vector obtained in this way is combined with the drive signal vector obtained in the past sub-frame and then delayed by 1 sample by the delay circuit 150 for 20 to 147 samples, and K samples are obtained. It is stored in the adaptive codebook 110 in units. Next, a configuration for sequentially correcting the drive signal vector in the noise codebook 120, which is the gist of the present invention, by learning will be described. Figure 1
In, a training vector creation unit 162 and a learning unit 163 are provided for this learning.

When the search for the drive signal vector from the noise codebook 120 ends in a certain subframe, the optimum drive signal vector N _opt is output from the noise codebook 120. The training vector creation unit 162 sets this drive signal vector as the training vector V _t . The learning unit 163 uses the training vector from the training creation unit 162 to generate the noise codebook 12
The drive signal vector stored in 0 is successively corrected by learning. This correction is performed in parallel with the encoding process.

FIG. 2 shows the procedure of this learning. First, the training vector V _t from the training vector creation unit 162 is input (S1). Next, among the plurality of drive signal vectors stored in the noise codebook 120, a vector to be corrected (updated) is set (update area setting S2). As a method of setting the update area, a method of setting a representative vector existing within a certain Euclidean distance from the training vector V _t in the update area is used. Here, the drive signal vector in the noise codebook is paraphrased as a representative vector. Further, the size of the update area is assumed to decrease with time. Letting NE (i) be the update area at time i, NE (i) has the following properties.

[0041]

[Equation 8]

Next, the representative vector in the update area is updated (corrected) using the training vector V _t . Representative vector V _j (i) included in the update area at time i
Is updated according to the following equation:

[0043]

[Equation 9] Where α (i) is a variable that controls the magnitude of the correction,
It has the following properties.

[0044]

[Equation 10]

Then, the above update is continued until it is determined whether or not the update has converged (S4). The convergence is determined depending on whether the following expression is satisfied, and if it is satisfied, it is determined that the convergence is achieved.

[0046]

[Equation 11]

This learning method is one of the learning methods of the neural network known as the Kohonen algorithm. Regarding this Kohonen algorithm, for example, T. Kohonen's Self-Organization an
Since it is described in d Associative Memory, Springer-Verlag (1984) (Reference 3), detailed description is omitted. The learning method is not limited to this, and other learning methods may be used.

By such learning, the drive signal vector in the noise codebook 120 has a property that is statistically similar to the drive signal vector used as the training vector. As described above, the drive signal of the synthesis filter is created so that the error between the input audio signal to be encoded and the synthesis signal is minimized. Therefore, learning is performed using this drive signal, and the noise codebook 12
By modifying the drive signal vector in 0, a noise codebook suitable for generating a synthesized voice with a small difference from the input voice, that is, a low distortion, will be created.

Moreover, since the learning is performed in parallel with the voice encoding process, the property of the drive signal vector in the noise codebook 120 also changes in response to the change in the property of the input voice signal. As a result, it is possible to synthesize high-quality speech at a low bit rate such as an encoding rate of 8 kbps or less and even when the number of bits assigned to the drive signal is small.

In other words, in the conventional CELP system, the same noise codebook is always used to reproduce the voice signal regardless of the change in the characteristics of the input voice signal. On the other hand, in this embodiment, by the learning operation as described above,
The drive signal vector in the noise codebook changes so that the error of the synthesized signal with respect to the input speech signal becomes smaller. As a result, if the number of bits assigned to the drive signal is the same, higher quality synthetic speech can be obtained.

The coding parameters obtained in the above process are multiplexed by the multiplexer 142 and sent from the output terminal 143 to the transmission path as a coded output. That is, in the multiplexer 142, the LPC analysis circuit 10
The information of the LPC prediction coefficient obtained in 2 is used as the encoding circuit 10
3, the code of the index of the adaptive codebook 110 obtained by the minimum distortion search circuit 115, the code of gain information multiplied by the multiplier 111 by the gain encoding circuit 140, and the minimum The code of the index of the noise codebook 120 obtained by the distortion search circuit 125 and the code obtained by encoding the gain information multiplied by the multiplier 121 by the gain encoding circuit 141 are multiplexed. Next, the configuration of a speech decoding apparatus corresponding to the speech encoding apparatus of FIG. 1 will be described with reference to FIG.

In FIG. 3, the input coding parameters are first decomposed into individual parameters by the demultiplexer 201, and then decoded by the decoders 202, 203 and 204, respectively. Then, a drive signal is created based on the decoded adaptive codebook index and gain, and the noise codebook index and gain. This drive signal is filtered by the synthesis filter 215 to create a synthetic voice signal.
The synthesized voice signal is output from the output terminal 217 after the spectrum is shaped by the post filter 216 and the auditory distortion is suppressed.

In FIG. 3, the noise codebook 2
A training vector creation unit 262 and a learning unit 263 are provided for learning the drive signal vector in the 12. These have the same functions as the training vector creation unit 162 and the learning unit 163 in the speech coding apparatus shown in FIG. 1, respectively, and their operations are also the same, so detailed description will be omitted.

As is clear from the present embodiment, in the present invention,
The signal used for training is set to the signal obtained by both encoding and decoding. As a result, because of the learning of the codebook, it is not necessary to transmit any auxiliary information and the bit rate does not increase. Next, FIG. 4 shows a block diagram of a learning-type speech encoding apparatus according to the second embodiment of the present invention.

In the first embodiment, the contents of the noise codebook are updated by learning, but the contents of the adaptive codebook may be updated. The present embodiment is an example of the configuration for learning the adaptive codebook. In FIG. 4, a buffer 131,
A training vector creation unit 132, a learning unit 133, a memory 134, and a delay circuit 135 are provided.

When the search for the drive signal vector from the adaptive codebook 110 and the vector from the noise codebook 120 is completed in a certain subframe, the adder 130 outputs a drive signal vector for a new synthesis filter. The buffer 131 adds this new drive signal vector to the drive signal vector of the past sub-frame and stores it. Specifically, as shown in FIG. 5, the buffer 131 is composed of a shift register having an accumulated data length of M _B samples, and a total of M _B samples including the drive signal vector newly output from the adder 130. The information of the drive signal vector of is accumulated. The information on the drive signal vector in the buffer 131 is read by the training vector creation unit 132. As shown in FIG. 5, the training vector creation unit 132 sequentially sets the information of the drive signal vector from the buffer 131 by sequentially setting the length of the vector dimension K as one unit.
Clipping is performed while shifting each sample, and this is sent to the learning unit 133 as a training vector. In FIG. 5, m =
Although it is 1, values such as m = 2 and 3 may be used. Further, in FIG. 5, M _B = 2K. For example, m = 1, M
_{When B} = 2K, K-1 is used as the training vector.
Vectors will be created.

The learning unit 133 sequentially corrects the drive signal vector stored in the adaptive codebook 110 by learning using the training vector from the training vector creation unit 132. This correction is performed in parallel with the encoding process.

FIG. 6 shows the procedure of this learning. First, a training vector is input from the training vector creation unit 132 (S1). Next, of the plurality of drive signal vectors stored in the memory 134, the vector having the maximum similarity to the input training vector is searched (S2). The reciprocal of the Euclidean distance can be used as the similarity. The drive signal vector in the memory 134 is stored in the shift register as a signal sequence having a length N as shown in FIG. The drive signal vector is generated by cutting out while shifting the sample from the right end of the shift register to the left one sample at a time with the length of the vector dimension K as one unit. Let n be the total number of drive signal vectors in the adaptive codebook.

[0059]

[Equation 12] Have a relationship. Next, the similarity vector C _j obtained in step S2 is updated as follows using the training vector V _t (S3).

[0060]

[Equation 13]

Here, α is a coefficient for controlling the weight of the weighted average of C _j and V _t , and can be a predetermined constant or a value that adaptively changes according to the similarity. The drive signal vector of the memory 134 is updated by the above equation, but in reality, a part of the signal series in the shift register from which the drive signal vector C _j is cut out is updated.
By repeating the above processing until it is determined in S4 that the training vector is exhausted, the memory 13
The drive signal vector in 4 is learned. When this learning is completed, the signal series stored in the shift register of the memory 134 is stored in the cutout adaptive codebook 110 while shifting by one sample by the delay circuit 135 with the length of the dimension K of the drive signal vector as one unit. To do. This completes the learning of the adaptive codebook. It is not necessary to actually prepare the adaptive codebook, and the memory 134
Can be virtually an adaptive codebook.

By such learning, the drive signal vector in the adaptive codebook 110 has a property statistically similar to the drive signal vector used as the training vector. Moreover, since the learning is performed in parallel with the voice coding process, the property of the drive signal vector in the adaptive codebook 110 also changes in response to the change in the property of the input voice signal. As a result, the coding rate is 8k
Even if the number of bits allocated to drive signal encoding is small at a low bit rate such as bps or less, it is possible to synthesize high-quality speech.

In the conventional CELP system, when the nature of the input voice signal suddenly changes from unvoiced sound to voiced sound,
Since the content of the adaptive codebook is only the drive signal vector in the unvoiced section, it is not possible to immediately generate the periodic drive signal necessary for synthesizing the voiced sound, and it is difficult to follow the change of the input voice signal. Become. As a result, there is a problem that the clarity of the synthesized voice is deteriorated. On the other hand, in the present embodiment, even when the input voice signal suddenly changes from unvoiced sound to voiced sound, the driving signal vector of the past voiced sound section is stored in the adaptive codebook by the learning operation described above. , It is possible to synthesize voiced sound by using this drive signal vector, and it becomes possible to obtain clear synthesized speech.
Further, the drive signal vectors in this embodiment have a relationship of overlapping with each other as apparent from FIG. 7, and the amount of calculation required to search for the optimum drive signal vector from the adaptive codebook can be reduced. The conventional adaptive codebook also has a structure in which each vector overlaps as described in Document 2, and an optimum drive signal vector can be searched efficiently. In the present embodiment, even if the contents of the adaptive codebook are randomly updated by the learning operation, the overlap structure is not broken, and it is possible to efficiently search the drive signal vector. Since an efficient search method using the overlap structure is described in Document 2, it is omitted here. The encoding parameters obtained in the above process are multiplexed by the multiplexer 142 and output to the output terminal 14
3 is transmitted to the transmission path as encoded output.

The configuration of the speech decoding apparatus corresponding to the speech encoding apparatus of FIG. 4 is as shown in FIG. In FIG. 8, a memory 224 and a delay circuit 225 are provided for learning the drive signal vector in the adaptive codebook 210. These have the same functions as the memory 134 and the delay circuit 135 in the speech coding apparatus shown in FIG. 4, respectively, and their operations are also the same, so detailed description will be omitted.

[0065]

As described above, according to the present invention, the driving signal vector in the adaptive codebook and the noise codebook has the same statistical property as the driving signal used as the training vector. On the other hand, the drive signal of the synthesis filter refers to the input speech signal to be encoded, and the optimum drive signal vector from the adaptive codebook and the noise codebook, that is, the error between the input speech signal and the synthesis speech signal by the synthesis filter is It is created by searching for a drive signal vector that minimizes. Therefore, it is possible to create synthetic speech with less distortion with respect to the input speech signal by sequentially correcting the driving signal vectors in the adaptive codebook and the noise codebook by learning using the optimum driving signal vector. An adaptive codebook and a noise codebook suitable for Further, since the learning process itself can proceed in parallel with the encoding process, the properties of the adaptive codebook and the noise codebook also change in response to the changes in the properties of the input speech signal.

As a result, the present invention can be applied even at a low bit rate of about 8 kbps or less, which is difficult to secure the quality due to the limitation of the number of bits allocated to the drive signal in the conventional method that does not perform the learning as described above. According to this, it becomes possible to synthesize high-quality speech. Moreover, since the training signal for learning is set to the driving signal vector that can be obtained by both the encoding and decoding processes, there is no need to transmit any auxiliary information for learning, and there is no increase in bit rate. ..

[Brief description of drawings]

FIG. 1 is a block diagram of a learning-type speech encoding apparatus according to a first embodiment of the present invention.

FIG. 2 is a diagram for explaining a procedure of learning a drive signal vector in the embodiment.

FIG. 3 is a block diagram of a speech decoding apparatus according to the embodiment.

FIG. 4 is a block diagram of a learning-type speech encoding apparatus according to a second embodiment of the present invention.

FIG. 5 is a diagram for explaining a method of creating a training vector in the same embodiment.

FIG. 6 is a diagram for explaining a procedure for learning a drive signal vector in the embodiment.

FIG. 7 is a diagram showing how drive signal vectors are stored in the memory according to the embodiment.

FIG. 8 is a block diagram of a speech decoding apparatus according to the embodiment.

FIG. 9 is a block diagram showing a configuration related to a drive signal vector search in a conventional speech encoding device.

[Explanation of symbols]

100 ... Audio signal input terminal 102 ... LPC
Analysis circuit 103 ... Encoding circuit 106 ... Weighting filter 107 ... Weighting synthesis filter 110 ... Adaptive codebook 112 ... Weighting synthesis filter 114 ... Square error calculation circuit 115 ... Minimum distortion search circuit 120 ... Noise codebook 122 ... Weighting synthesis filter 124 ... Square error calculation circuit 125 ... Minimum distortion search circuit 131 ... Buffer 132 ... Training vector creation unit 133 ... Learning unit 134 ... Memory 135 ... Delay circuit 140 ... Gain coding circuit 141 ... Gain coding circuit 142 ... Multiplexer 143 ... Output Terminal 150 ... Delay circuit 162 ... Training vector creation unit 163 ... Learning unit

Claims (3)

[Claims] 1. A codebook in which a drive signal vector is stored as a codeword, a search means for searching an optimum drive signal vector from the codebook by referring to an input voice signal, and an optimum search searched by the search means. A synthesis filter for synthesizing a voice signal using a drive signal vector, a training vector creation means for creating a training vector using the optimal drive signal vector, and a training vector created by this means in the codebook And a learning unit that sequentially corrects the driving signal vector of 1. 2. A plurality of codebooks in which driving signal vectors are stored as codewords, a search means for searching an optimum codeword from the plurality of codebooks by referring to an input voice signal, and a search by the search means. A synthesis filter for synthesizing a speech signal using the optimized codeword as a driving signal vector, a training vector creating means for creating a training vector using the optimal codeword, and a training vector created by this means A learning-type speech coding apparatus, comprising: learning means for sequentially correcting at least one corresponding codeword in the codebook. 3. A plurality of codebooks in which drive signal vectors are stored as codewords, a search means for searching an optimum codebook from the plurality of codebooks by referring to an input voice signal, and a search by the search means. A synthesis filter for synthesizing a voice signal by using the optimized codeword as a drive signal vector, a training vector creation means for creating a training vector using the drive signal vector obtained from the optimum codeword, and this means A learning-type speech encoding apparatus, comprising: a learning unit that sequentially corrects at least one corresponding codeword in the codebook by using the training vector created by the above.