CN102063897B - A sound bank compression and usage method for embedded speech synthesis system - Google Patents

Abstract

The invention discloses a sound library compression and use method for an embedded voice synthesis system, which is used for an embedded operating system and converting received random characters into voice to be output. Syllables in Chinese are used as basic units of a synthesis system and a voice model library; firstly, an original voice model base based on syllables is established, then data decomposition and compression are carried out on the original voice model base, and finally a final compression model base is obtained through data recombination. The method provided by the invention can reduce the space resource occupied by the synthesis system under the embedded platform, improve the synthesis speed and simultaneously keep good naturalness and tone quality of the synthesized voice.

Description

A sound bank compression and usage method for embedded speech synthesis system

technical field

The present invention generally relates to a method for compressing and using a sound bank for an embedded speech synthesis system, especially for terminal equipment with limited storage and computing resources.

Background technique

The purpose of speech synthesis technology is to allow machines to restore natural human speech. Embedded devices are widely used, and terminal embedded devices frequently interact with users. Voice is the most natural means of interaction. A general speech synthesis system can be divided into three main functional modules: text analysis module, prosody generation module and acoustic synthesis module. The splicing and synthesis method based on large-scale corpus is widely used because of its simple technology and high quality of synthesized sound. However, this method has a large sound bank, and although the space can be reduced by clustering, encoding, and compression, the sound quality is damaged and the flexibility is reduced. Therefore, in recent years, the statistical modeling parameter synthesis method based on large-scale corpus has been widely studied. The basic idea is to perform parametric representation and statistical modeling on a large number of original speech databases, and select models according to specific rules to form a model sequence during synthesis. The parameter sequence of the synthesized sentence is calculated, and the speech meeting the requirements is synthesized through the method of parametric synthesis. The speech synthesized by parametric statistical modeling method has high degree of naturalness and intelligence. In this method, in order to ensure the synthesis effect, the original speech library needs to cover prosodic features as much as possible, and the obtained model library can reach hundreds of megabytes. After model clustering, the model library can be compressed to about ten megabytes. This scale can meet the storage and computing capabilities of some mid-to-high-end devices such as handheld computers, but it still cannot meet the practical requirements for terminal devices with limited computing and storage resources.

In the training process of parametric statistical modeling speech library, the speech feature parameters often used are pitch frequency, spectral coefficient and duration feature, and the parametric model is Hidden Semi-Markov Model (HSMM). According to the state transition characteristics of the Hidden Semi-Markov Model (HSMM), the model of each feature includes a decision tree of each state and a probability distribution function representing the leaf nodes of the decision tree. At present, the commonly used probability density function representation method is a single Gaussian model. In the finally obtained model, the model of spectral coefficient occupies 80% to 90% of the space of the final model size, which is the part that needs to be compressed most. The existing methods to reduce the size of the spectral parameter model adopt methods such as reducing the numerical precision, controlling the clustering factor and bundling the variance. On the premise of using syllables as the basic unit of the synthesis system, when controlling the amount of training data to the minimum value acceptable for the hearing of synthesized speech, the model library obtained based on the above method also requires at least 1 megabyte of storage space. Also, with tighter controls on clustering, the naturalness and quality of the synthesized speech can be significantly degraded. The above-mentioned system is still expensive for devices with limited resources, and it is difficult to meet the needs of users. Therefore, there is a need for an improved method for realizing a parametric speech synthesis system that occupies less resources under an embedded platform.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for compressing and using a sound bank applied to an embedded Chinese speech synthesis system. It enables the speech model library to occupy a very small space resource, improves the computing speed, and maintains good synthesis naturalness and sound quality.

To achieve the above objectives, this paper provides a method for compressing and using parametric statistical models, which is used to reduce the space occupied by the model library and maintain the sound quality of the synthesis. The training and synthesis process of the original model base uses Chinese syllables as the basic unit; the compression process of the model base is divided into the following three steps:

A. Create a library of primitive models based on Chinese syllables.

B. Decompose the single Gaussian distribution representing the original spectral model into three parts: energy, spectral mean and spectral variance. The spectral mean and spectral variance are compressed separately by vector quantization technology.

C. Combining the energy, the compressed spectral mean codebook, the index, and the global variance to obtain the final compressed model library.

The method for compressing and using the above-mentioned parameterized statistical model is characterized in that: the creation process of the original model library based on Chinese syllables as a unit is divided into the following five steps:

A. Create an original phonetic library based on Chinese syllables.

B. Extract pitch frequency, spectrum parameters and duration parameters of all syllables in the speech library. Train a syllable model that does not take context into account.

C. Train a syllable model that considers contextual context based on the contextual information of all syllables, and use a decision tree-based method to perform state clustering on the model.

D. Further train the clustered model parameters.

E. Return to step C and repeat steps C and D to output a parameterized statistical model.

The method for compressing and using the above-mentioned parameterized statistical model is characterized in that: the process of compressing the spectrum model is divided into the following six steps:

A. Divide the state Gaussian distribution of the spectral model into three parts: energy, spectral mean and spectral variance. This method considers first-order dynamic features and second-order dynamic features.

B. Use the mean vectors of all state distributions (including static features, first-order dynamic features and second-order dynamic features) as training samples for codebook training of vector quantization.

C. Search for the training sample with the smallest distance from the codebook in each class after vector quantization classification, and save it instead of the codebook of this class.

D. Reclassify the training samples with the new codebook.

E. Judging whether the new classification result is the same as the original classification result. If yes, the spectral mean vector quantization codebook training ends; if no, return to step C and repeat steps C and D.

F. Average the variance vectors (including static features, first-order dynamic features and second-order dynamic features) of all state distributions to obtain a global variance vector.

The above method for compressing and using the parameterized statistical model is characterized in that: the model recombination process is as follows: the state distribution in the original model is replaced by the energy value and the corresponding mean value vector codebook index, and finally stored in the global variance value.

The above method can greatly compress the spectral model using syllables as primitives, while maintaining the sound quality and naturalness of the original model synthesis.

In order to better meet the requirements of the computing speed of embedded devices, the present invention also provides an embedded speech synthesis system. Include the following four steps:

A. Text analysis and prosody generation module, which is used to analyze the content of the synthesized text to obtain the corresponding syllable sequence. At the same time, each syllable is attached with relevant prosody information marked by context and context, and its format is the same as that used for model training. ;

B. The model decision module is used to receive the above-mentioned syllable sequence attached to the prosodic information, use the model decision tree obtained through training to generate a corresponding model state sequence, and obtain the duration decision result;

C. The parameter sequence generation module is used to receive the above-mentioned model state sequence, utilizes the described compressed spectrum model to calculate the global variance windowing matrix, and finally calculates the spectral parameter sequence and the fundamental frequency parameter sequence;

D. The voice waveform synthesis output module is used to receive the parameter sequence, generate the voice waveform data to be synthesized, and output it for playback or storage.

Above-mentioned embedded speech synthesis system is characterized in that: described parameter sequence generation module is divided into following 5 steps:

A. Calculate the energy sequence and the spectral coefficient sequence according to the state sequence of the spectral coefficient, and obtain the pitch frequency sequence according to the state sequence of the pitch frequency;

B. Calculate the global variance matrix from the global variance. In the process of parameter generation, the feature parameters to be synthesized are calculated by dimension-by-dimensional generation, and the mean value or global variance of one dimension is taken for each calculation;

C. Obtain a one-dimensional spectrum mean codebook sequence according to the state spectrum mean codebook sequence corresponding to the received model state sequence;

D. Solve the characteristic parameter sequence according to the received global variance matrix and the state spectrum mean value codebook sequence;

E. Judging whether all spectral coefficients have been processed. If yes, the spectral coefficient calculation ends; if not, return to step C and repeat steps C and D.

The embedded speech synthesis system established according to the above method can be completely applied in the embedded system, and the occupied space resource and required calculation complexity are not more than the capability of the embedded device.

The present invention will be further described below in conjunction with the accompanying drawings and embodiments, and the steps and processes for realizing the present invention will be better described by referring to the detailed description of each component of the system in conjunction with the accompanying drawings.

Description of drawings

Accompanying drawing 1 is based on the structural block diagram of the embedded speech synthesis system of Chinese syllable

Figure 2 Schematic diagram of the compression process of the spectrum model

Figure 3 Schematic diagram of the generation process of spectrum parameters

In the figure 1. Text input, 2. Text analysis and prosody generation, 3. Model decision-making, 4. Parameter generation, 5. Waveform synthesis, 6. Voice output, 7. Training voice library, 8. HMM model training, 9. Data Decomposition, 10. Model compression, 11. Data reorganization, 12. Compression model library, 13. Whether the codebook is stable, 14. Data reclassification, 15. Codebook search and replacement, 16. Vector quantization, 17. State spectrum mean, 18 . State spectrum variance, 19. Variance average, 20. End, 21. Yes, 22. No, 23. Start, 24. Calculate the global variance matrix, 25. Obtain the codebook sequence of the corresponding dimension of the state, 26. Solve the one-dimensional spectrum Coefficient sequence, 27. Whether the 18-dimensional parameters have been processed, 28. End, 101. Model training part, 102. Speech synthesis system.

Detailed ways

In accompanying drawing 1, in the embodiment of the present invention, speech synthesis system of the present invention is deployed in a kind of embedded operating system, and this embedded speech synthesis system comprises: model training part (101), speech synthesis system (102 ).

Among them, the model training part (101) is only used offline in the system, and is only used to generate the compressed model library (12) required for the speech synthesis system to work. Wherein the training voice bank (7) includes the original voice recorded, and the process of generating the compressed model bank (12) offline by the training voice bank (7) includes: HMM model training step (8), data decomposition (9), model compression (10) and model reorganization (11).

In the HMM model training step (8), firstly, the recorded original training speech library is automatically segmented in units of syllables by using the Speech Recognition Toolkit (HTK), to obtain rough cutting boundary information, and manually proofread. Then mark according to the context and prosodic information of the syllable, including: the current syllable, the current syllable tone, the previous syllable, the previous syllable tone, the next syllable, the next syllable tone, and the high-level prosody obtained by grammatical analysis of the text Information, that is, the position and number of low-level prosodic units in high-level prosodic units. The prosodic levels considered in this example include prosodic words, prosodic phrases and sentences. Use the speech synthesis model training toolkit (HTS) to train the HSMM model on the original training speech, and the model content includes pitch frequency, spectral coefficient and duration parameters. The duration is represented by the number of frames, and the frame length is 5 milliseconds. The model is divided into 10 states, and each state uses a single Gaussian model to represent the state probability distribution. During the training process, the size of the pitch frequency and duration model can be controlled by appropriately controlling the model clustering factor to obtain the original speech model library.

In data decomposition (9), the state Gaussian distribution of the spectral model is divided into three parts: energy, spectral mean and spectral variance. All features considered in this example consider a combined feature consisting of its static, first-order dynamic, and second-order dynamic features. The spectral coefficients used in this example are 18-dimensional line spectral pair (LSP) coefficients, and the state of the HSMM model is represented by a single Gaussian distribution. Therefore, each original spectrum model state contains a 57-dimensional mean vector and a 57-dimensional variance vector. Each decomposed original spectrum model data is expressed as a 3-dimensional energy vector, a 54-dimensional mean vector, and a 57-dimensional variance vector. All the original state spectrum means (17) constitute mean value quantization codebook training data, and all original state spectrum variances (18) constitute global variance training data.

In model compression (10), as shown in accompanying drawing 2. The model compression process is divided into the following five steps:

Vector quantization (16), using the mean value vector as the vector quantization codebook training data, and using the LBG algorithm to train an initial codebook. During the training process, according to the characteristics of the selected 18-dimensional line spectrum pair (LSP) coefficients, the distance between the coefficients is weighted, and the weighted line spectrum pair distance between the two line spectrum pair coefficients x and y is defined as:

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 54</span>}}<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 18</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 18</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&Delta;x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&Delta;y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 18</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>$

where the weighting coefficients are:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; )</span>$

xn, xn+1, xn-1, yn, yn+1, yn-1 are the n-th dimension, n+1-dimensional and n-1-dimensional static coefficients of features x and y, respectively. Δx _n , Δy _n , Δ ² x _n , Δ ² y _n are the corresponding first-order dynamic coefficients and second-order dynamic coefficients.

The codebook search replaces (15), searches for the training sample with the smallest distance from the codebook in each class after the vector quantization classification, and replaces the codebook of this class and saves it. The distance judgment criterion adopts the weighted line spectrum pair distance.

Data reclassification (14), using a new codebook to reclassify the training samples, and the classification distance judgment criterion adopts the weighted line spectrum pair distance.

Whether the codebook is stable (13), judge whether the new classification result is the same as the original classification result. If yes, the spectral mean value vector quantization codebook training ends; if not, return to data reclassification (14) and repeat data reclassification (14).

Variance averaging (19), average the variance vectors (including static features, first-order dynamic features and second-order dynamic features) of all state distributions to obtain the global variance vector.

In data reorganization (11), the state distribution in the original model is replaced by the energy value and the corresponding mean value vector codebook index, and finally stored in the global variance value. The rest of the models are combined sequentially as required to form a compressed model (12). So far, the model training part (101) of the speech synthesis system has finished its work.

As shown in accompanying drawing 1, text input (1) receives the input text, and in the embodiment of the present invention, the system provides the input interface that can be used for handwriting input or text selection and pasting.

The speech synthesis system (102) further includes text analysis and prosody generation (2), model decision-making (3), parameter generation (4) and waveform synthesis (5). The text analysis and prosody generation module converts the received input Chinese character text string into a syllable string with relevant contextual information attached. The model decision module determines the duration of each state from the received syllable string according to the model decision tree obtained through training, and obtains the model state sequence of pitch frequency and spectral coefficient. The parameter generation module calculates the energy sequence and the spectral coefficient sequence according to the state sequence of the spectral coefficient, and obtains the pitch frequency sequence according to the state sequence of the pitch frequency. As shown in Figure 3, the calculation process of the spectral coefficient sequence is divided into the following four steps:

Calculate the global variance matrix (24), and calculate the global variance matrix according to the global variance. In the process of parameter generation, the characteristic parameters to be synthesized are calculated in a dimension-by-dimensional way, and the mean value or global variance of one dimension is taken for each calculation. There are two types of global variance matrices used, namely:

_WUWn ＝W× _Un ×W

and

WU _n =W×U _n

Where U is the nth-dimensional global variance and its first-order and second-order dynamic coefficients, W is the dynamic window coefficient, and the multiplication is matrix multiplication;

Obtain the codebook sequence (25) of the corresponding dimension of the state, and obtain the one-dimensional spectrum average codebook sequence according to the state spectrum mean codebook sequence corresponding to the received model state sequence;

Solve the one-dimensional spectral coefficient sequence (26), and solve the characteristic parameter sequence according to the received global variance matrix and state spectrum mean codebook sequence, the method is to solve the following matrix equation:

${\mathrm{<span\; class="notranslate">\; WUW</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; wu</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}$

Wherein L is the desired line spectrum pair coefficient sequence, and the equation solving method can adopt any linear matrix equation solving algorithm that meets the computing requirements of the embedded system. In this example, the solution method of LU decomposition and forward and backward Gaussian elimination is adopted for the coefficient matrix;

Judging whether all 18-dimensional spectral coefficients have been processed (27). If yes, then the spectral coefficient solution ends; if not, return to obtain the codebook sequence (25) of the corresponding dimension of the state, repeat the acquisition of the codebook sequence (25) of the corresponding dimension of the state, and solve the one-dimensional spectral coefficient sequence (26).

So far, the work of the parameter generation module is over. The adoption of the global variance matrix scheme provided by the present invention can improve the operation speed and save the space consumption of intermediate steps.

In the waveform synthesis step, any algorithm that can meet the resource requirements of the device can be used, for example, the synthesis filtering method adopted in G.723, or the synthesis filtering method in other speech decoding algorithms based on linear prediction.

Voice output (6), used for playing or storing synthesized digital voice signals.

The invention relates to a method for compressing and using a sound bank applied to an embedded Chinese speech synthesis system. Based on this method, the speech model library can occupy a very small space resource, improve the calculation speed, and maintain a good synthetic naturalness and sound quality.

When the present invention is used on an embedded device, all audio input and output can use the input and output interfaces provided by the device itself. Voice can be turned on or off on the device at any time. When the voice function is not enabled, various functions of the original device will not be affected in any way.

The above examples are preferred embodiments of the present invention, and the application of the present invention can be used in various embedded terminal devices. According to the main concept of the present invention, those skilled in the art can generate various low or equivalent applications. Therefore, the protection of the present invention should be based on the protection scope of the claims.

Claims (3)

1. A compression method for a parameterized statistical model of an embedded speech synthesis system, which is used to reduce the space occupied by the model library and maintain the sound quality of the synthesis; the training and synthesis process of the original model library uses syllables in Chinese as the basic unit; the method steps are as follows: A. Create an original model library based on Chinese syllables; B. Decompose the original spectral model into three parts: energy, spectral mean and spectral variance; use vector quantization technology to compress the spectral mean and spectral variance respectively; Step B is divided into: B1. Divide the state Gaussian distribution of the spectral model into three parts: energy, spectral mean and spectral variance; B2. the mean vector of all state distributions is used as a training sample, and the codebook training of vector quantization is carried out; the mean vector of all state distributions includes static features, first-order dynamic features and second-order dynamic features; B3. Search for the training sample with the smallest distance from the codebook in each class after vector quantization classification, and save it instead of the codebook of this class; B4. Reclassify the training samples with a new codebook; B5. Judging whether the new classification result is the same as the original classification result, if yes, the spectral mean value vector quantization codebook training ends; if not, return to step B3, and repeat steps B3 and B4; B6. The variance vectors of all state distributions are averaged to obtain a global variance vector; the variance vectors of all state distributions include static features, first-order dynamic features and second-order dynamic features; C. Combining energy, compressed spectral mean codebook and index and global variance to obtain a compressed spectral model, and merging it with other models to obtain a final compressed model library, said other models including duration and pitch frequency models. the 2. the compression method of parameterized statistical model according to claim 1 is characterized in that step A is divided into: A1. Create an original voice library based on Chinese syllables; A2. Extract the pitch frequency, spectral parameters and duration parameters of all syllables in the speech library; train the syllable model without considering the context; A3. According to the context information of all syllables, train the syllable model considering the context, and use the method based on decision tree to cluster the state of the model; A4. Further train the clustered model parameters; A5. Return to step A3, repeat steps A3 and A4, and output the parameterized statistical model. the 3. the compression method of parameterized statistical model according to claim 1 is characterized in that in the step C, the model recombination process is as follows: the state distribution in the original spectrum model is replaced with energy value and corresponding mean value vector code book index, Finally, the global variance value is stored; other models are stored sequentially as needed. the

Images