RU2343564C2 - Method of voice signal variable-structure system-based adaptive encoding - Google Patents

Abstract

FIELD: physics, communication.

SUBSTANCE: invention relates to system of telecommunication and is intended for coding voice signals based on variable-structure system. Proposed method of encoding comprises separating the input voice signal segments into six classes, i.e. a pause, tone segment, noise segment of the first type, noise segment of the second type, transition segment of the first type, transition segment of the second type, and encoding of the input voice signal recognised segments by various methods, varying the coding system structure.

EFFECT: improved quality of synthesised voice signal at fixed low transmission rate.

2 dwg

Description

The present invention is intended for encoding speech signals (PC) based on a system with a variable structure, the use of which is aimed at reducing the redundancy of transmitted information.

Known methods for encoding speech signals based on linear prediction and various excitation signals of a synthesizing filter using the vector quantization of excitation signals and parameters describing the spectral envelope of a speech signal, for example [1, 2].

The disadvantage of such methods is the low ability to adapt these algorithms to the properties of the processed PC, which determines the insufficient quality of signal recovery at the reception. In these algorithms, only the encoder parameters are changed during the encoding process, and its structure remains unchanged. The fixed structuring of the space of encoded parameters and the constancy of the powers of the subspaces of representations (the prediction order for linear prediction, the size of code books for vector quantization, the length of the encoded vector) inherent in existing algorithms and expressed in a fixed code structure do not create the necessary conditions for maximizing the use of a priori information about speech signal, which prevents further optimization of the codec.

A known method of encoding speech signals based on linear prediction depending on the type of processed segment of the speech signal [3], where the quality of the synthesized signal is improved by using the classification of the processed frames of the speech signal into two disjoint classes: voiced and unvoiced speech, and coding of segments related to different classes by various methods. The disadvantages of this method include a small number of classes into which the speech signal is subdivided, adaptive redistribution of the power of the subspaces of representation of the encoded parameters under the conditions of a fixed structure of the encoding device, which determines the insufficient quality of signal recovery at the reception. Given the existing requirements for digital representation of speech signals, more careful processing of the speech signal is required. Under these conditions, methods based on changing only the parameters of the encoder in accordance with the characteristics of speech become unacceptable and do not provide sufficient quality when encoding a PC.

The proposed method of speech conversion solves the problem of improving the quality of the synthesized PC without increasing the transmission speed.

The specified technical result is achieved in that in real time the input speech signal is divided along the time axis into segments, the input speech segment is recognized as a pause, tone segment, noise segment of the first type, noise segment of the second type, transition segment of the first type, transition segment of the second type based on the following classification procedure (figure 1).

At the first stage of classification, the signal is divided into active sections and pauses, the criterion for decision making is the ratio:

where N is the number of samples in the processed speech segment;

s _i - counting of a speech signal;

P ₀ is the threshold value of the power characteristic determined experimentally.

In the case of inequality (1), it is decided that the processed segment 1 belongs to the class of pauses 2. Otherwise, a decision is made whether the processed speech segment belongs to the class of active segments 3.

At the second stage of classification, active speech segments 3 are divided into 4 types of segments: tone 7, noise 4, transitional first type 5 and transitional second type 6. For this division, the tone / noise (TS) parameter and the fundamental frequency (OT) F _from on the analysis segment. The calculation of the TS and F signals _{from is} carried out jointly based on the analysis of the PC autocorrelation function (ACF) and the Takura-Saito method. Using two methods together reduces the likelihood of errors in the classification of speech segments. In this case, the decision rules on the type of segment are formulated as follows.

Tone segments 7 include segments for which:

- ACF analysis method defines the segment as tonal;

- The Takura-Saito method defines the segment as tonic.

Noise segments 4 include segments for which:

- ACF analysis method determines the segment as noise;

- The Takura-Saito method defines the segment as noise.

The transitional segments of the first type 5 include segments for which:

- ACF analysis method determines the segment as noise;

- The Takura-Saito method defines the segment as tonic.

The transitional segments of the second type 6 include segments for which:

- ACF analysis method defines the segment as tonal;

- The Takura-Saito method defines the segment as noise.

At the third stage of classification, the noise segments of speech 4 are divided by the envelope coefficient and the power characteristic of the signal (1) into two classes. The decision rule is determined by the ratio:

where P - is determined in accordance with the left side of the expression (1);

α ₀ - threshold value of the coefficient of complexity of the encoded segment, determined experimentally;

η is the envelope coefficient of the encoded signal, defined as:

If, as a result of the calculation, inequality (2) is satisfied, a decision is made that the segment being processed belongs to the noise segment of the first type 8, otherwise, to the noise segment of the second type 9.

Then, a segment of the input speech signal is encoded by encoding the waveform if the segment of the input speech signal is defined as a pause, a noise segment of the first type or a noise segment of the second type, or the remnants of short-term predictions of the input speech signal are found and the remnants of short-term predictions are used using sinusoidal analytical coding if a segment of the input speech signal is defined as a tone segment, a transition segment of the first type, or a transition segment of the second type.

Thus, in accordance with the obtained statistical and parametric characteristics, the encoding device structure (blocks 10 or 11) is selected that provides a minimum of distortion of the speech signal.

In the drawings (Figs. 1 and 2) the essence of the proposed solution is presented, in which Fig. 1 shows a classification of recognizable segments of speech in accordance with the proposed solution, Fig. 2 is a structural diagram of a speech encoding device based on a system with a variable structure.

The proposed method of converting a speech signal can be implemented in a device for encoding speech signals (figure 2).

The original speech signal is supplied to the PCM encoder 12, which implements the conversion of the analog signal into digital form according to ITU Recommendation G.711. In the block for the formation and initial processing of the analysis segment PC 13, the digitized speech signal is segmented into identical subframes equal to the quasistationary period. Next, the subframes of the speech signal are sequentially fed to the speech / pause analyzer 14, the statistical and parametric characteristics extractor 15, the subframe forming unit 24 and the linear predictor 26 structure and parameters control unit. In the speech / pause analyzer 14, speech is divided into activity and pause segments, when of this, the speech segments assigned to active are transmitted for subsequent analysis to the tone / noise analysis unit 18, and control signals about the decision made (speech / pause) from this block are transferred to the stat selector -terrorist and parametric characteristics of the structure 15 and the control subsystem 17. The codec unit 15 is realized the selection and parametric statistical characteristics of the speech signal segment in allocating it to the active speech segments. The subframe generating unit 14 is designed to separate the subframes of the vector quantization procedure 30 on the analysis segment, the results of the procedure are supplied to the structure control unit of the vector quantizer 25 and the vector quantizer 30. In the tone / noise analysis unit 18, the tone-noise signal is extracted on the analysis segment, when assigned its block 14 to the segment of active speech. In this case, in the case of a noise signal, a control signal that carries information about this decision is sent to block 17, in the opposite case (tone signal extraction), the control signal is sent to block 19.

In blocks 17 and 19, the encoder structure control subsystem is implemented, while block 17 controls the encoder structure depending on classification decisions relating the processed speech segment to pause and noise segments, and block 19 uses information from block 18 about the activity and tonality of the speech segment. Information signals from block 19 are fed to the OT frequency separator based on the analysis of ACF 20 and the OT frequency separator by the Takura-Saito method 21. In the indicated blocks 20 and 21, the fundamental frequency is selected based on the analysis of the autocorrelation function of the speech signal analysis segment and using the Takura method -Saito respectively. The calculation results are sent to the OT 22 frequency value adjustment block, in which the fundamental tone frequency value is corrected to make decisions about the type of speech segment being processed by the speech frame classifier 16 and select optimal operating modes by the structure control unit and parameters of the linear predictor 26 and the vector structure control unit quantizer 25. Thus, at the inputs of the classifier 16 receives information signals: from the output of the statistical isolator and parameter of the physical characteristics of PC 15 and the output of the unit for adjusting the frequency from OT 22. The classification results by block 16 are sent to the encoder control subsystem 23, which determines the encoding mode depending on the segment classification result, the output of this block is control signals for the subframe generation unit PC 14, structure control unit vector quantizer 25 and the control unit structure and parameters of the linear predictor 26.

In accordance with the classification results, block 25 controls the operation of the vector quantizer 30, as well as selects codebooks of various structures 27 that most closely match the encoded speech subframe. Block 26 controls the structure and parameters of the linear predictor. The operation of the linear predictor is associated with the code books of the parameters of the short-term linear predictor 31 and the code books of the parameters of the long-term linear prediction 32, with the block for calculating the parameters of the short-term linear prediction 28 and the block for calculating the parameters of the long-term linear prediction 29, with the block of short-term linear analysis 33 and the block of long-term linear analysis 34 in which linear prediction procedures are directly implemented based on the linear pre sayings selected from the corresponding code books and the most appropriate calculated. Block 26 also interacts with a block for selecting the best structure and parameters of a linear predictor 36, which selects the best structure and parameters of a linear predictor based on the analysis procedure through synthesis and the results of control actions on the structure of the encoder. Information signals obtained as a result of vector quantization (block 30) and / or linear prediction (block 36) are fed to the input of the output sequence generating unit of the encoder 35, which implements the transmission frame of the encoder.

The decoding procedure on the receiving side consists in extracting from the received sequence of the transmission frame information about the type of structure and parameters of the encoded PC, selecting the appropriate decoder structure and setting the PC according to the received excitation signal and the parameters of the synthesizing device.

The above data show that the introduction of the procedure for classifying speech segments into 6 types into the coding system: pause, tone segment, noise segment of the first type, noise segment of the second type, transition segment of the first type, transition segment of the second type, and coding of recognized segments of the input speech signal by different methods by changing the structure of the coding system can improve the quality of the synthesized PC without increasing the transmission speed.

Information sources

1. Ustinov A.A., Tyulegenev A.O., Danilyuk V.V. Patent No. 2152646, cl. 7 G10L 21/00. The method of compression and restoration of speech signals. Bull. No. 19 dated 10.07.2000.

2. Kostrov V.V., Dyranov Yu.V., Factory S.Yu. Patent No. 2166804, cl. 7 G10L 13/02. A method of converting speech and a device for its implementation. Bull. No. 13 dated 05/10/2001.

3. Nishiguchi M., Iijima K., Matsumoto D., Omori S. Patent No. 2233010, class. 7 G10L 19/06. Methods and devices for encoding and decoding speech signals. Bull. No. 20 dated July 20, 2004.

Claims (1)

A method of adaptive coding of speech signals based on a system with a variable structure, namely, that the input speech signal is divided into segments along the time axis, the remnants of short-term predictions of the input speech signal are found, the input speech signal is recognized as voiced or unvoiced, and the residual short-term predictions are encoded using sinusoidal analytic coding, if part of the input speech signal is defined as voiced, or the input speech signal is encoded n with a waveform coding, if a part of the input speech signal is determined to be non-voiced, characterized in that the segments of the input speech signal are recognized as pause, tone segment, noise segment of the first type, noise segment of the second type, transition segment of the first type, transition segment of the second type, then encode the segment of the input speech signal by encoding the waveform if the segment of the input speech signal is defined as a pause, noise segment of the first type or noise segment of the second type, or residuals of short-term predictions of the input speech signal and encode residuals of short-term predictions using sinusoidal analytical coding if the segment of the input speech signal is defined as a tone segment, a transition segment of the first type, or a transition segment of the second type.

Images