JPH08147000A - System for encoding and decoding by vector quantization - Google Patents

Abstract

PURPOSE: To prevent an error in vector quantization from increasing extremely even in the case where a speaker is changed in a learning vector quantization which updates the contents of a code book so that they are specialized to be adapted to a specific speaker. CONSTITUTION: The system for encoding and decoding by the learning vector quantization which adaptively updates the contents of the code book 6 on the basis of the result of matching between a vector to be quantized and a code vector detects a change of feature parameters of a speech signal by a speaker change detection part 12 and judges that the speaker changes when the change is detected, thereby setting an initialization flag. A rewriting control part 8 initializes the contents of the code book 6 with an initialization code book 7 holding the contents of the contents of the code book 6 in its initial state once the initialization flag is set. Consequently, the contents of the code book after the speaker changes are initialized into averaged contents for an unspecified speaker and the vector quantization error is prevented from increasing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an encoding / decoding system by vector quantization for highly efficiently compressing voices, images and the like for transmission or recording using a coatbook.

[0002]

2. Description of the Related Art Conventionally, vector quantization has been known for compressing voices and images with high efficiency for transmission or recording. In vector quantization, a signal waveform to be quantized is cut out for a certain section, and the waveform pattern of this section is expressed by one code (index). Therefore, various waveform patterns are stored in advance and an index is added to each waveform pattern. Various waveform patterns are vector information expressed by the number of dimensions of the number of samples. These vector information are called code vectors, and the correspondence table between these code vectors and indexes is called a codebook.

[0003] For example, CELP ("Code-Excited Linear Prediction:") which is known as one of audio coding systems.
High Quality Speech at Very low Bit Rate "MRSch
Roeder, BSAtal ICASSP, 1985) divides a prediction residual component of speech into data of about 40 samples called subframes, and applies vector quantization to each subframe. However, the codebook used for this vector quantization is not optimized for any particular individual as it will be average of many speakers. For this reason, the error between the quantized vector and the codebook vector is large, and good quality speech cannot be transmitted. If the number of code vectors stored as a codebook is large, the sound quality is improved, but on the other hand, the amount of codebook is large and the cost is increased, and the information amount of the index is also increased, so that the transmission amount is also increased. There is also a problem that it takes time to search the codebook.

Therefore, a learning vector quantizer (LVQ) that suppresses the occurrence of quantization distortion with a small amount of codebook information is disclosed in Japanese Patent Laid-Open No. 2-1886836. This learning vector quantizer also transmits the quantized vector to the decoding device side when the minimum distortion at the time of vector quantization is larger than a predetermined value, and selects the earliest selected quantum in the codebook. By rewriting the coded representative vector (code vector), the contents of the codebook are adaptively updated according to the characteristics of the quantized vector.

[0005]

However, in the above-mentioned conventional coding / decoding apparatus for learning vector quantization, the original codebook created so as to be evenly matched with many average speakers is specified. It is configured to be updated sequentially so as to be specialized for the speaker. Therefore, when the speaker changes, there is a problem in that the codebook specialized for the speaker before the replacement is damaged, the vector quantization error increases, and the decoded speech deteriorates.

The present invention has been made in view of such a problem, and in the learning vector quantization for updating the contents of the codebook so as to be specialized for the speaker, even when the speaker changes. An object of the present invention is to provide a coding / decoding method by vector quantization in which the vector quantization error does not increase extremely.

[0007]

A coding / decoding method by vector quantization according to the present invention includes a code vector indicating some typical signal patterns in a predetermined section and an index for specifying each of the code vectors. A codebook storing the correspondence table of is provided in both the encoding device and the decoding device, and in the encoding device, the codebook that best matches the quantized vector to be quantized based on the encoding target signal , And outputs the index, and in the decoding device,
While decoding the quantized vector by obtaining the code vector corresponding to the index obtained on the encoding device side with reference to the codebook, the quantized vector and the code in the encoding device In a coding / decoding method by vector quantization provided in both the coding apparatus and the decoding apparatus, codebook rewriting means for adaptively updating the contents of the codebook based on the result of matching with a vector, in the initial state, An initialization codebook holding the contents of the codebook is provided in both the encoding device and the decoding device, and the encoding device extracts the characteristic parameter of the signal to be encoded, and the feature extraction means. And a feature change detecting means for detecting a change in the feature parameter extracted by the means. Means, when said change in said characteristic parameters characteristic change detecting means is detected, characterized in that said that the contents of the codebook intended to initialize the contents of said initial codebook.

Further, according to a further preferred aspect of the present invention, further comprising a temporary storage means for temporarily storing the contents and characteristic parameters of the codebook prior to the initialization of the codebook, the codebook rewriting means. When a change in the characteristic parameter is detected by the characteristic change detection means, first, the characteristic parameter extracted by the characteristic extraction means is compared with the characteristic parameter stored in the temporary storage means, and both are compared. If they are similar, the contents of the codebook are rewritten to the contents of the codebook stored in the temporary storage means, and if the feature parameters are not similar, the contents of the codebook are temporarily changed. The contents are stored in the storage means and then initialized to the contents of the initialization codebook.

The encoding device supplies initialization information indicating that the contents of the codebook are rewritten to the contents of the initialization codebook to the decoding device, and the decoding device supplies the supplied initial information. The contents of the codebook may be initialized to the contents of the initialization codebook based on the encoding information, and the encoding device once decodes the encoding result to be output, and the feature extraction is performed on the decoding result. Extracting the characteristic parameters by means, and initializing the contents of the codebook to the contents of the initialization codebook when a change in the extracted characteristic parameters is detected. Decoding the encoded result,
A feature parameter may be extracted by the feature extracting means with respect to this decoding result, and the contents of the codebook may be initialized to the contents of the initialization codebook when a change in the extracted feature parameter is detected. Good.

When the signal to be encoded is a voice signal, the feature change detection means changes the presence / absence of voice, the pitch change, and the feature coefficient of the voice signal output from the feature extraction means. , A change in spectrum envelope information, a change in gain, a change in voiced / silent of the voice signal output in the process of encoding in the encoding device, a change in pitch, a change in characteristic coefficient, a change in gain, And the codebook rewriting means may be controlled based on at least one of the changes in the vector quantization error.

[0011]

According to the present invention, in the encoding / decoding system by the learning vector quantization for adaptively updating the contents of the codebook based on the matching result of the vector to be quantized and the code vector, the characteristics of the signal to be encoded are Changes in parameters are detected, and if this change is detected, it is determined that the speaker has changed, and the contents of the codebook are initialized. Can be average content for unspecified speakers,
It is possible to prevent an increase in vector quantization error.

In a telephone, another speaker often answers the telephone for a little while, or another sound such as a melody is temporarily input. In such a case, it is not efficient to detect that the speaker has changed and initialize the codebook each time. Therefore, prior to the initialization of the codebook, a temporary storage means for temporarily storing the contents of the codebook and the characteristic parameters is provided, and when a change in the characteristic parameters is detected, first, the characteristic parameters and The characteristic parameters saved in the temporary storage means are compared. If they are close to each other, it is determined that the previous speaker has returned, and the contents of the codebook are rewritten to the contents stored in the temporary storage means. Initialize the codebook as if there were. As a result, when another speaker answers the telephone temporarily, the codebook in which the previous learning result is reflected can be used again, and accurate vector quantization can be achieved.

Further, the initialization of the codebook needs to be performed by both the encoding device and the decoding device, but by supplying the initialization information from the encoding device to the decoding device, it is synchronized with the encoding device. The codebook can be initialized in the decoding device. Furthermore, if the encoding device and the decoding device are set to perform the initialization of the codebook according to the same algorithm, the encoding device and the decoding device can be decoded without supplying the above initialization information to the decoding device side. The initialization timing can be synchronized with the device. For example, in the encoding device, the encoded result is once decoded (local decoding), and the feature is extracted from the decoded result to detect the change. The decoding device also decodes the received coding result, performs feature extraction on the decoding result, and detects the change. In this way, by applying the same algorithm to the same information, it is possible to perform a codebook initialization process that is synchronized between the encoding device and the decoding device.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a CELP according to an embodiment of the present invention.
FIG. 2 is a block diagram showing a speech signal coding apparatus including vector quantization according to the method, and FIG. 2 also showing a decoding apparatus.
In FIG. 1, a speech signal, which is a signal to be encoded and is input to the encoding device, is A / D converted by the A / D converter 1 and then supplied to the long-term predictor 2 and the feature extraction unit 11.
The long-term predictor 2 converts the input digital voice signal into
For example, it is divided into frames of 160 samples and analyzed for each frame, and long-term prediction coefficients such as percoll coefficients, gains, and pitches are output to the multiplexing unit 3. Further, the long-term predictor 2 subtracts the signal linearly predicted by the long-term prediction coefficient from the input digital audio signal to obtain a long-term prediction residual component, and outputs this to the frame division unit 4. The frame division unit 4 uses the frame timer to
The long-term prediction residual component of 60 samples is divided into, for example, 40-sample subframes, and is output to the vector quantization unit 5 as a 40-dimensional prediction residual vector. In CELP, this prediction residual vector becomes a quantized vector to be vector-quantized. The vector quantizer 5 normalizes the prediction residual vector so that the magnitude of the vector becomes 1, and searches the codebook 6 for the code vector closest in distance to the normalized prediction residual vector. , The index IDX and the normalization coefficient are output to the multiplexing unit 3.

Further, in addition to the codebook 6, this coding apparatus is provided with an initialization codebook 7 which holds the contents when the codebook 6 is initialized. The contents of the codebook 6 are sequentially updated by the rewriting control unit 8. Therefore, the codebook 6 is a rewritable R
It is composed of a memory such as AM. On the other hand, since the contents of the initialization codebook 7 are fixed, the ROM
And the like.

An error between the prediction residual vector input to the vector quantizer 5 and the code vector selected by vector quantization is calculated by the subtractor 9. The error vector from the subtractor 9 is given to the error amount detector 10. When the magnitude of the error vector is larger than the predetermined value, the error amount detection unit 10 sets the error transmission flag “1” and supplies this flag and the error vector as update information to the rewrite control unit 8 and the multiplexing unit 3. To do. As a result, the rewrite control unit 8 is controlled and the content of the codebook 6 is updated. Further, when the magnitude of the error vector is smaller than the predetermined amount, the error amount detection unit 10 sets the error transmission flag to “0” and supplies it to the rewrite control unit 8 and the multiplexing unit 3.

Further, the voice signal supplied to the voice feature extraction unit 11 is extracted with a feature parameter by the voice feature extraction unit 11. This characteristic parameter is information peculiar to the speaker such as a change in voiced / non-voiced sound signal, a change in pitch, a change in characteristic coefficient, a change in spectrum envelope information, and a change in gain. These characteristic parameters are supplied to the speaker change detection unit 12. When the change in the characteristic parameter is detected, the speaker change detection unit 12 recognizes that the speaker has changed and sets the initialization flag to "1". This initialization flag is also supplied to the rewrite control unit 8, and the rewrite control unit 8 initializes the contents of the codebook 6 to the contents of the initialization codebook 7 when the initialization flag is “1”. Multiplexing unit 3
Encodes and multiplexes the long-term prediction coefficient, IDX, its normalization coefficient, update information, and the initialization flag, and transmits them to the transmission path or records them on a recording medium.

On the other hand, on the decoding device side, as shown in FIG. 2, the received or read signal is supplied to the demultiplexing unit 21 and is decoded / demultiplexed to obtain the long-term prediction coefficient and It is divided into a normalization coefficient, an index IDX, update information, and an initialization flag. The long-term prediction coefficient and the normalization coefficient are supplied to the long-term prediction synthesis unit 22. Further, the decoding device side is also provided with the codebook 24 and the initialization codebook 25 having the same contents as the encoding device, and the vector dequantization unit 23 refers to the codebook 24 from the supplied index IDX. By
Select a code vector. Codebook 2
The code vector output from 4 is the long-term prediction synthesis unit 2
2 is supplied. The long-term prediction synthesizing unit 22 restores the code vector to the original signal level by the normalization coefficient, and synthesizes this signal and the signal obtained by linear prediction from the long-term prediction coefficient. This synthesized signal is D / A converted by the D / A converter 26 and output as a decoded audio signal. Also,
The update information and the initialization flag separated by the demultiplexing unit 21 are supplied to the rewriting control unit 27. Rewriting control unit 27
Has the same function as the rewrite control unit 8 of the encoding device, updates the contents of the codebook 24 based on the update information, and initializes the contents of the codebook 24 based on the initialization flag. To the contents of.

Next, the operation of the encoding / decoding system configured as described above will be described. When voice is input,
The characteristic part, that is, the Percoll coefficient, the pitch, the gain, etc. are extracted by the long-term predicting unit 2 and the feature extracting unit 11, but the remaining residual components are It will be encoded and decoded. FIG. 3 is a diagram showing the contents of the code books 6 and 24. At the time of initialization, the codebooks 6 and 24 include, for example, 512 typical prediction residual components selected from various languages, races, genders, etc., so as to be compatible with all unspecified speakers. Code vectors are stored in advance.
Further, the codebooks 6 and 24 store the index IDX for identifying the code vector in a form corresponding to each code vector. In addition to this, a flag indicating whether or not each code vector can be rewritten, information regarding weighting of selection priority of each code vector, and the like may be added.

If the number of code vectors is large, the quantization error is reduced, and high quality encoding / decoding can be realized, but on the other hand, the memory capacity must be increased and the cost is increased. , Search also takes time. Further, for example, if there are 512 types of code vectors, 9-bit information may be transmitted as the index IDX, but if the number of types of code vectors increases, the number of bits to be transmitted also increases.

When the pattern vectors forming the codebooks 6 and 24 are expressed as a two-dimensional vector on a two-dimensional plane for the sake of simplicity, a black circle in FIG. 4A is obtained. Here, the numerical value written in black circles is the index of each code vector. When a vector Vi indicated by "x" in the figure is input to the vector quantizer 5 as a prediction residual vector to be quantized, this vector Vi is converted into the vector V12.
Since the index 12 is closest to, the index 12 is selected as the quantization result in the vector quantization unit 5.

Since the tendency of the residual component input to the vector quantizer 5 is peculiar to each speaker, the codebooks 6 and 2 are used.
In No. 4, it is necessary to update the internal code vector to the best contents by learning. Therefore, when the error amount detection unit 10 of the encoding device detects that the quantization error for a specific code vector continuously exceeds a certain value, the code vector selected by the codebook 6 and The contents of the codebook 6 are updated by replacing the original code vector with a vector obtained by adding the average value of the quantization error to a new code vector. Also,
At this time, the average value of the quantization error is transferred to the decoding device side together with the error transmission flag “1”, and the codebook 24
Update the contents of. FIG. 4B shows an example in which the code vector V12 shown in FIG. 4A is replaced with the code vector newly generated from the quantization error. At this time, a new code vector may be replaced with the code vector with the lowest selection frequency per predetermined time. In this case, a flag that prohibits replacement may be set for the primitive vector existing at the outer edge of the vector space. Codebook 2
The updating process of No. 4 may be a process of changing the weighting of each code vector without rewriting such a code vector.

By updating the codebooks 6 and 24 as described above, the contents of the codebook can be adaptively changed according to the characteristics of the speaker. However, when the power is turned on or the speaker changes, the characteristics of voice input can be changed. In case of sudden change, the contents of the learned codebook may be rather inconvenient. In this case, it is desirable to initialize the contents of the codebooks 6 and 24 to the original contents. Therefore, the voice feature extraction unit 11 extracts the speaker's feature parameter from the voice signal, and the speaker change detection unit 12 detects that the speaker's feature parameter has changed. The following are possible speaker characteristic parameters.

(A) Sound / Silence The ratio of sound / silence differs for each speaker. Further, when the silent portion occurs longer than the average included in the general speaking language, the speaker is likely to change. (B) Voice Pitch Since the voice pitch is the vibration frequency of the vocal cords, and by extension, the voice pitch, the frequency distribution (time average, variance, etc.) reflects individual characteristics. (C) Speech Feature Coefficients As speech feature coefficients, there is a Percoll coefficient (reflection coefficient) that represents the vocal tract with an all-pole filter, but the time average, variance, etc. thereof differ from speaker to speaker. (D) Spectrum envelope The frequency of the voice is analyzed and the shape of the spectrum envelope is
The temporal change rate and the change variance also have the same tendency as in (c). (E) Voice gain The voice gain varies depending on the habit of the speaker. Further, the rate of change (variance) of the gain also changes according to the method of applying intonation, etc., so that it can be used for individual discrimination. (F) Vector quantization error When the codebook specialized for the first speaker is completed, if the speaker changes, the quantization error naturally increases. Therefore, the speaker change can also be discriminated by monitoring the change in the quantization error.

The above-mentioned characteristic parameters may be those extracted by the characteristic parameter extraction unit 11 or those obtained in the process of vector quantization. For example,
In CELP encoding, pitch, Percoll coefficient, gain, voiced / non-voiced flag, etc. are generated in the CELP algorithm, which can be used. In addition, in a coding / decoding method of a format that performs spectrum analysis and compresses information using the envelope information,
It is possible to use the spectrum envelope parameter of (d) as it is.

These parameters are used as the speaker change detection unit 12
, The speaker change detection unit 12 firstly temporarily increases the quantization error and the like due to the way the speaker speaks,
For example, preprocessing such as LPF (low-pass filter) is applied to the input parameters so that a temporary change in pitch, characteristic coefficient, etc., such as when a whistle is blown, is not detected as a change in the speaker. Then, for example, the total evaluation value etotal is obtained based on the following determination formula.

[0027]

[Equation 1] e total = ka Pa + kb Pb + kc Pc + k
d Pd + ke Pe + kf Pf Pa to Pf: Each parameter ka to kf: Weighting coefficient for evaluation of each parameter

The total evaluation value eto thus obtained
After tal is further subjected to LPF as post-processing, it is compared with a predetermined threshold value, and if it exceeds the threshold value, it can be determined that the speaker has changed.

When the speaker change detecting unit 12 detects a speaker change, the initialization flag becomes "1", and the rewriting control unit 8 changes the contents of the codebook 6 to the contents of the initialization codebook 7. initialize. Since the initialization flag is transmitted or recorded, the same initialization of the codebook 24 is performed also in the decoding device. At this time, all the contents of the codebooks 6 and 24 may be rewritten, but considering the efficiency of rewriting, for example, a primitive code vector that does not contain much personal information is
The amount of data to be rewritten may be reduced by adding a rewrite prohibition flag or the like. Further, together with the initialization flag, characteristic parameters such as pitch information are supplied to the rewriting control unit 8, and male / female voice discrimination is performed by referring to these parameters. Then, if the initial value of the codebook is initialized with a typical example of male voice or a typical example of female voice based on the determination result, the vector quantization error after the speaker change can be reduced.

As described above, the initialization codebooks 7 and 25 may be configured by ROM so as to store all the initialization data, but the initialization codebooks 7 and 25 are rewritten. It may be configured with a possible memory so that only the rewritten vector information is stored.
Even if all the initialization data are stored in the initialization codebooks 7 and 25, the rewriting of the codebooks 6 and 24 is performed only on the rewritten data, so that the rewriting time can be shortened. Good.

FIG. 5 is a block diagram showing an encoding apparatus in an encoding / decoding system according to another embodiment of the present invention. The system of this embodiment is different from the system of the above-described embodiment in that a temporary storage memory 31 for temporarily saving the codebook and the characteristic parameters and a timer 32 for measuring the saving time are provided.

In other words, on the telephone, the speaker may change for a little while, saying "I'm a little different". Therefore, when the initialization flag becomes “1”, the rewrite control unit 8 saves the characteristic parameters of the previous speaker and the contents of the codebook 6 before the rewriting in the temporary storage memory 31, and then sets the timer 32. It starts and initializes the contents of the codebook 6. The timer 32 is set to, for example, several minutes, and when the initialization flag becomes “1” before the timeout occurs, the feature parameter saved in the temporary storage memory 31 and the feature parameter extracted this time When the two are close to each other, it is determined that the previous speaker has been returned, and the contents of the codebook saved in the temporary storage memory 31 are stored in the codebook 6. Write back to. If a time-out occurs when the speaker is switched, the probability that the previous speaker speaks is considered to be low, and the contents and characteristic parameters of the codebook 6 are saved, and then the codebook 6 is initialized. Further, when the speaker is switched, if the feature parameter of this time and the saved feature parameter are not close to each other, it is determined that another speaker has been completely replaced, and the contents of the codebook 6 and After saving the characteristic parameters, the codebook 6 is initialized. Whether the parameters are close to each other can be determined by using the following discriminant equation.

[0033]

[Equation 2] dperson = √ [ka '(Pma-Pa) ² + kb' (Pmb-Pb) ² + ......... + kf '(Pmf-Pf) ² ] dperson: Evaluation value Pma of whether the parameters are close to each other -Pmf: Parameters temporarily stored Pa-Pf: Parameters obtained after the speaker is switched ka'-kf ': Judgment weighting coefficient for each parameter

Regarding the parameters Pma to Pmf and Pa to Pf, it is desirable to carry out preprocessing such as LPF also in this case. Codebooks 6 and 2 of the encoding device and the decoding device
In order to achieve the consistency of 4, the restoration flag indicating that the contents of the temporary storage memory 31 have been restored may be transmitted or recorded from the encoding device in addition to the initialization flag. The decoding device restores or initializes the contents of the codebook 24 based on the restoration flag and the initialization flag.

In the above embodiment, the initialization flag and the restoration flag are supplied from the encoding device to the decoding device. However, the encoding device and the decoding device may generate these pieces of information by using exactly the same algorithm. For example, the transfer of such information becomes unnecessary. 6 and 7 are diagrams showing an example of such a system, FIG. 6 is a block diagram of an encoding device, and FIG. 7 is a block diagram of a decoding device. For example, when an analysis and synthesis method such as the CELP method is used, synthetic speech is temporarily created inside the encoding device. Hereinafter, this synthesized voice is referred to as locally decoded data. When the long-term prediction unit 2 obtains a prediction residual vector, locally decoded data is obtained inside the long-term prediction unit 2. This locally decoded data matches the decoded data synthesized by the decoding process of the decoding device. Therefore, in the encoding device, the speech feature extraction unit 11 extracts a feature parameter from the locally decoded data, and the extracted feature parameter and the feature coefficient, pitch, etc. obtained by the long-term prediction unit 2 are used as a speaker change detection unit. 12 to detect changes in the characteristic parameters. The delay unit 41 is a delay unit for compensating for a delay in processing in vector quantization.

On the other hand, also in the decoding device, the voice feature extraction unit 51 extracts feature parameters from the decoded data synthesized by the long-term prediction synthesis unit 22, and the speaker feature detection unit 52 uses the extracted feature parameters. And the received or read characteristic coefficient, pitch, and other information to the speaker change detection unit 52 via the delay unit 53. As a result, the speaker change can be detected with the same data and the same algorithm as the encoding device, so that it is not necessary to supply an initialization flag between the encoding device and the decoding device. Therefore, data transmission or storage capacity can be reduced.

In the above embodiments, the vector quantization of the prediction residual vector of the voice signal is taken as an example, but the same applies to the vector quantization of the voice signal itself or the image signal. Needless to say.

[0038]

As described above, according to the present invention,
In a coding / decoding method by learning vector quantization that adaptively updates the contents of the codebook based on the matching result between the quantized vector and the code vector, the change in the characteristic parameter of the signal to be coded is detected, and this change is detected. If is detected, it is determined that the speaker has changed, and the contents of the codebook are initialized.Therefore, the contents of the codebook after the speaker change is averaged for unspecified speakers. It is possible to prevent the increase of vector quantization error.

[Brief description of drawings]

FIG. 1 is a block diagram of an encoding device by vector quantization according to an embodiment of the present invention.

FIG. 2 is a block diagram of a decoding device using the same vector quantization.

FIG. 3 is a diagram for explaining the contents of a codebook in the encoding device and the decoding device.

FIG. 4 is a diagram schematically showing the contents of code vectors before and after updating the codebook.

FIG. 5 is a block diagram of an encoding device by vector quantization according to another embodiment of the present invention.

FIG. 6 is a block diagram of an encoding apparatus by vector quantization according to still another embodiment of the present invention.

FIG. 7 is a block diagram of the decoding device.

[Explanation of symbols]

1 ... A / D converter, 2 ... Long-term prediction unit, 3 ... Multiplexing unit, 4
... Frame division section, 5 ... Vector quantization section, 6, 24 ...
Codebook, 7, 25 ... Initialization codebook, 8, 2
7 ... Rewrite control unit, 9 ... Subtractor, 10 ... Error amount detection unit, 1
1, 51 ... Voice feature extraction unit, 12, 52 ... Speaker change detection unit, 21 ... Demultiplexing unit, 22 ... Long-term prediction synthesis unit, 23 ...
Vector dequantization unit, 26 ... D / A converter, 31 ... Temporary storage memory, 32 ... Timer, 41, 53 ... Delay.

Claims (7)

[Claims] 1. A codebook, which stores a correspondence table of code vectors indicating some typical signal patterns in a predetermined section and indexes for specifying the code vectors, an encoding device and a decoding device. Provided in both, in the encoding device, select the code vector in the codebook that best matches the quantized vector to be quantized based on the encoding target signal and output the index, in the decoding device, While decoding the quantized vector by obtaining the code vector corresponding to the index obtained on the encoding device side with reference to the codebook, the quantized vector and the code in the encoding device A codebook book that adaptively updates the contents of the codebook based on the result of matching with a vector In an encoding / decoding method by vector quantization in which a conversion unit is provided in both the encoding device and the decoding device, an initialization codebook holding the contents of the codebook in the initial state is provided in both the encoding device and the decoding device. In addition to the above, the encoding device includes a feature extraction unit that extracts a feature parameter of the encoding target signal, and a feature change detection unit that detects a change in the feature parameter extracted by the feature extraction unit, The codebook rewriting means initializes the content of the codebook to the content of the initialization codebook when the characteristic change detecting means detects a change in the characteristic parameter. Encoding and decoding method by vector quantization. 2. The method further comprises: a temporary storage unit for temporarily storing the contents and characteristic parameters of the codebook prior to the initialization of the codebook, wherein the codebook rewriting unit is the characteristic change detecting unit. When a change in the parameters is detected, first, the characteristic parameters extracted by the characteristic extracting means are compared with the characteristic parameters stored in the temporary storage means, and when the two are approximate, the code The contents of the book are rewritten to the contents of the codebook stored in the temporary storage means, and when the two characteristic parameters are not close to each other, the contents of the codebook are saved in the temporary storage means and then the initialization is performed. The encoding / decoding method by vector quantization according to claim 1, wherein the encoding / decoding method is initialized to the contents of a codebook. 3. The encoding device supplies initialization information indicating that the content of the codebook has been rewritten with the content of the initialization codebook to the decoding device, and the decoding device provides the supplied information. 3. The coding / decoding method by vector quantization according to claim 1, wherein the content of the codebook is initialized to the content of the initialized codebook based on the initialization information. 4. The encoding device temporarily decodes the output encoding result, and the feature extracting means extracts the feature parameter from the decoded result, and a change in the extracted feature parameter is detected. The content of the codebook is initialized to the content of the initialization codebook when the decoding device decodes the input coding result, and the decoding result is characterized by the feature extraction means. 3. The parameter extraction is executed, and when the change of the extracted characteristic parameter is detected, the contents of the codebook are initialized to the contents of the initialization codebook.
An encoding / decoding method by the described vector quantization. 5. The encoding target signal is a voice signal, and the feature change detection means changes voiced / unvoiced changes, pitch changes, and feature coefficient changes of the voice signal output from the feature extraction means. , A change in spectrum envelope information, a change in gain, a change in voiced / silent of the voice signal output in the process of encoding in the encoding device, a change in pitch, a change in characteristic coefficient, a change in gain, And controlling the codebook rewriting means on the basis of at least one of the change of the vector quantization error.
An encoding / decoding method by vector quantization according to any one of 1. 6. A codebook, which stores a correspondence table of code vectors showing some typical signal patterns in a predetermined section and indexes for specifying those code vectors, and a codebook based on a signal to be coded. Vector quantizing means for selecting the code vector in the code book that most matches the quantized vector to be quantized and outputting the index, and based on the matching result of the quantized vector and the code vector Codebook rewriting means for adaptively updating the contents of the codebook, an initialization codebook holding the contents of the codebook in the initial state, and a feature extracting means for extracting feature parameters of the signal to be coded, A characteristic change detecting means for detecting a change in the characteristic parameter extracted by the characteristic extracting means, The codebook rewriting means initializes the content of the codebook to the content of the initialization codebook when a change in the characteristic parameter is detected by the characteristic change detecting means. An encoding device using vector quantization. 7. A codebook that stores a correspondence table of code vectors indicating some typical signal patterns in a predetermined section and indexes that respectively specify those code vectors, and a codebook that corresponds to received indexes. Vector dequantization means for decoding the quantized vector by obtaining the code vector by referring to the codebook, and codebook rewriting means for updating the content of the codebook based on the received update information, An initialization codebook holding the contents of the codebook in an initial state, wherein the codebook rewriting means is based on an initialization flag received based on a change in a characteristic parameter of a signal to be coded. Is initialized to the contents of the initialization codebook. Decoder according Le quantization.