RU2445719C2 - Method of enhancing synthesised speech perception when performing analysis through synthesis in linear predictive vocoders - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: in the method of enhancing perception of synthesised speech when performing analysis through synthesis in linear predictive vocoders, iterative synthesis of a speech signal is performed at the transmitting side for duration of the speech quasi-stationary section, while changing codec parameters during each iteration in accordance with the linear prediction algorithm used for calculating the best linear predictive codec parameters within the given limits. Iterative calculation of the criterion of modified distortion of the spectrum of automatic adjustment and control units is carried out, while changing the codec parameters during each iteration in accordance with linear prediction algorithm used. Further, the minimum perception distortion is calculated, where codec parameters corresponding to this iteration are considered the best and are used for speech signal synthesis for the duration of the speech quasi-stationary section at the receiving side.

EFFECT: improved perception of synthesised speech when performing analysis through synthesis in linear predictive vocoders.

2 dwg

Description

The invention relates to the field of digital communication and can be used in telecommunication communication systems when implementing the analysis procedure through synthesis in vocoders with linear prediction.

Currently, there is a widespread use of information technology in telecommunication networks. The transition to digital signal processing and packet data transmission has allowed us to provide users with a wider range of information and communication services. Moreover, a rather large part of teletraffic in various applications is the transmission of speech signals.

The main problem of the digital representation of the speech signal is the task of high-quality and compact data coding for their transmission via digital communication channels. The solution to this problem will allow, under the conditions of a specified communication quality criterion, to increase the throughput of linear paths and transmission channels. Often, in some tasks of encoding a speech signal, it is supposed to reduce the transmission rate while maintaining qualitative indicators of its perception. In codecs of speech signals with a variable transmission rate, oriented to use based on the principle of packet switching in communication systems, it is appropriate to talk about a decrease in the average transmission rate while maintaining the quality of the synthesized speech signal.

Among the variety of methods for encoding speech signals, one of the most effective is the linear prediction method. The linear speech prediction method belongs to the class of methods using the speech signal model in the form of a response of a linear system with variable parameters (voice path) to the corresponding excitation signal (generating signal). The analyzer of the speech-converting device extracts from the short segment of the speech signal the state parameters of the linear system and the excitation signal, allowing the synthesizer to restore the original signal with the required degree of fidelity. To improve the quality of the synthesized speech signal in many linear-prediction-based speech coding algorithms, the excitation signal is complicated to represent in order to compactly transmit it to the receiving side, and on the other, to bring it closer to the form of prediction errors as an ideal signal for influencing the synthesis filter . In addition, different quantization options for various parameters of the lipreader (scalar, vector, and cascade vector) are introduced.

There are various algorithms for low-speed speech coding in vocoders with linear prediction (O.I. Shelukhin, N.F. Lukyantsev Digital processing and transmission of speech. - M .: Radio and Communication, 2000 - P.102-166; Rabiner L. R., Shafer R.V. Digital processing of speech signals. - M.: Radio and communications, 1981. - S.365-428). In many of them, one of the basic operations in low-speed coding is the analysis procedure through synthesis. A fairly detailed description of it is presented in (O.I.Shelukhin, N.F. Lukyantsev Digital processing and transmission of speech. - M: Radio and Communication, 2000 - S.108-112).

This procedure is iterative and is aimed at calculating the best linear codec parameters of the codec within the given constraints, and V is the number of possible iterations. At the same time, the synthesis of the speech signal and the further analysis of the suitability of the selected parameters are implemented in the encoder in order to transmit them via the communication channel to the receiving side, where the synthesis procedure will be implemented. Hence the name of this procedure.

Research in the field of speech coding indicated the need to use the perceptual characteristics of human hearing (Popov OB, Richter SG Digital signal processing in sound broadcasting paths. Textbook for universities. - M .: Hot line - Telecom, 2007. - S.13-21, 123-132). However, until now, as a criterion for choosing the parameters of a linear prediction codec, the analysis through synthesis uses either the standard deviation (1) or the total (2) or segmented (3) signal-to-noise ratio (SNR) based on the Euclidean metric and not taking into account the perceptual importance of the codec parameters in the synthesis of a speech signal (US Pat. Nos. 5,235,669 of 08/10/1993, RU No. 2248619 of March 20, 2005, RU No. 2331933 of August 20, 2008, RU No. 2343564 of January 10, 2009, US No. 6859775 of February 22. 2005).

where S (i) is the vector of input values;

S '(i) is the vector of estimates;

N is the total number of samples in the studied speech fragment.

, имеют равный вес. В общем случае для отражения вклада отдельных элементов в искажение вводятся неравные веса в виде взвешивающей матрицы. Указанный метод позволяет лишь сравнить форму огибающих исходного и синтезированного речевого сигнала. Поэтому для количественной оценки качества звучания синтезированного речевого сигнала во временной области чаще используют критерий отношения сигнал/шум:In expression (1), it is assumed that the distortions introduced by each element of the vector

have equal weight. In the general case, to reflect the contribution of individual elements to the distortion, unequal weights are introduced in the form of a weighting matrix. This method allows you to only compare the shape of the envelopes of the original and synthesized speech signal. Therefore, to quantify the sound quality of the synthesized speech signal in the time domain, the signal-to-noise ratio criterion is more often used:

S (i) and S '(i) -i-e samples of the original and synthesized speech;

N is the total number of samples in the studied speech fragment.

SNR takes into account the total signal and noise power over the entire duration of the test signal. In the study of some speech codecs, short-term signal-to-noise ratios calculated on short segments of the speech signal are of great importance. Thus, the segmented nature of the auditory perception of the elements of speech is taken into account. The criterion of the segment signal-to-noise ratio (SNR _seg ,) has the form:

where S (i) and S '(i) are the i-th samples of the initial and synthesized speech signals calculated on the m-th segment;

N is the length of the segment;

M is the number of segments in the speech fragment.

However, the above criteria of the objective estimation method reflect the noise level of the speech signal and show a weak correlation with the results of subjective tests when listening to speech segments. It should be noted that if the quality of the speech waveform encoders can be estimated by the degree to which the envelope shape of the reconstructed speech signal matches the original one using the above criteria, then for accurate low-speed parametric compression algorithms based on linear prediction, accurate reconstruction of the waveform is a difficult task. Therefore, methods for assessing the sound quality of a synthesized speech signal in the time domain are of little use. In order for the assessment of the sound quality of a speech signal to reflect the criteria of auditory perception, the principles of its formation should be based on the analysis of the spectral-correlation characteristics of speech.

The closest in technical essence to the claimed invention is US patent No. 6073092 from 06/06/2000, in which the calculation and minimization of the mean square error when implementing the analysis procedure through synthesis in a vocoder with linear prediction is performed.

The disadvantage of the used analysis procedure through synthesis is the mismatch of the human hearing system during the perception of synthesized speech and the proximity criteria used that determine the rules for analyzing the suitability of the selected codec parameters.

According to the known method for performing the analysis analysis through synthesis in a vocoder with linear prediction on the transmitting side, the speech signal is iteratively synthesized over the duration of the quasi-stationary part of speech, changing the codec parameters in accordance with the linear prediction algorithm used at each iteration, and the mean square error between the original and synthesized speech signal, find the iteration corresponding to the least standard error, with the parameter The codec's corresponding to this iteration are considered the best and based on them they form the codec's transmission frame and synthesize the speech signal over the duration of the speech quasistationary portion on the receiving side.

The objective of the invention, which determines its technical result, is to improve the perception of synthesized speech when implementing the analysis procedure through synthesis in vocoders with linear prediction.

This problem is solved in that in a method for improving the perception of synthesized speech when implementing the analysis procedure through synthesis in vocoders with linear prediction, the analysis and selection of the best parameters of the linear prediction codec within the given limits will be performed based on the calculation of estimates by the criterion of modified distortion of the spectrum of the MBSD barks ( Modified Bark Spectral Distortion) (4):

where M (n, i) and D (n, i) are the distortion values of the level of sensation and the difference in the intensity of sensation of the signal of the n-th speech segment in the i-th critical band;

N is the number of segments in the speech fragment;

K is the total number of critical bands.

This criterion is the most preferable, since it analyzes the spectral and correlation characteristics of speech, takes into account human hearing models, and it also shows a high correlation with the estimates obtained on the basis of subjective listening tests. Experimental studies have shown that in the case of the use of low-speed lipreder systems, the human auditory apparatus is more sensitive to frequency distortions arising from this than to amplitude and phase distortions (Pavlovets A.N., Petrovsky A.A. Use of the laws of psychoacoustics in the procedure of quantizing the parameters of the harmonic speech model signal // Speech technologies. 4, 2008, p. 55-60).

The calculation of the spectrum of barks is performed according to the expression (5):

where f is the frequency measured in Hertz;

b is the frequency measured in barges.

This issue is described in more detail in (Radzishevsky A.Yu. Fundamentals of analog and digital sound. - M .: Publishing house "Williams", 2006 - p.105-109). A detailed description of the MBSD criterion can be found in (W.Yang. Enhanced Modified Bark Spectral Distortion (EMBSD): An Objective Speech Quality Measure Based On Audible Distortion And Cognition Model / A Dissertation of the Requirement for the Degree Doctor of Philosophy - May, 1999. pp. 63-75).

According to this criterion, the synthesized and original speech signals on the quasistationary segment are divided into critical bands, in each of which the signal sensing intensity and the noise masking threshold are calculated, then the difference between the original and distorted sensation intensity values is determined in each band. If the obtained value of D (n, i) exceeds the calculated noise masking threshold NMT (n, i), then the sensation level M (n, i) is assigned the value 1, otherwise, the value 0.

The essence of the proposed method is as follows. When implementing the analysis through synthesis procedure in vocoders with linear prediction on the transmitting side, the speech signal is iteratively synthesized over the duration of the speech quasistationary section, changing the codec parameters at each iteration in accordance with the linear prediction algorithm used to calculate the best linear predictive codec parameters within the given constraints, they introduce an iterative calculation of the criterion for modified distortion of the spectrum of barks, changing the codec parameters at each iteration In accordance with the linear prediction algorithm used, then the minimal perceptual distortion is calculated, and the codec parameters corresponding to this iteration are considered to be the best and are used to synthesize a speech signal over the duration of the section of quasi-stationary speech on the receiving side.

The functioning algorithm of the proposed system that implements a method for improving the perception of synthesized speech when implementing the analysis procedure through synthesis in vocoders with linear prediction is presented in figure 1.

The analysis of the prior art made it possible to establish that analogues that are characterized by a combination of features that are identical to all the features of the claimed invention are absent, which indicates the compliance of the claimed method with the condition of patentability "novelty".

Thanks to a new set of essential features of the system, providing the introduction of an iterative calculation of the criterion of modified distortion of the spectrum of barks, calculating the minimum perceptual distortion and determining the best codec parameters for transmission over the communication channel in vocoders with linear prediction using the analysis through synthesis, significant improvement is achieved the quality of perception of synthesized speech on the receiving side.

Analysis of existing technical solutions in this and related fields of technology showed that the introduced distinguishing features are absent and do not follow explicitly from the prior art. The prior art also did not reveal the popularity of the impact provided by the essential features of the claimed invention, the transformations on the achievement of the specified technical result. Therefore, the claimed technical solution meets the criterion of "inventive step".

The claimed technical solution is illustrated by the drawing (figure 2), which shows a functional diagram of a device that implements a method for improving the perception of synthesized speech when implementing the analysis procedure through synthesis in vocoders with linear prediction.

A device that implements this method consists of a block for generating a segment of synthesized speech signal 1, to which the synthesized speech signal directly arrives. It is connected to the unit for calculating the values of the spectrum of barks in the critical bands of the segment of synthesized speech signal 3, in which the formation of a set of parameters describing the spectrum of barks occurs. The output of unit 3 is connected to the input of the unit for calculating the sensation intensity of the synthesized speech signal in the critical bands of segment 5. The original segment of the speech signal is fed to the unit for generating the segment of the original speech signal 2. Its output is connected to the input of the unit for calculating the spectrum of barks in the critical bands of the segment of the original speech signal four.

The formation of segments of the speech signal, carried out in blocks 1 and 2, is presented in (Bykov S.V., Zhuravlev V.I., Shalimov I.A.Digital telephony: Textbook for universities. - M .: Radio and communication, 2003 . - p. 66-72).

The calculation of the values of the spectrum of barks, carried out in blocks 3 and 4, is described in detail in (Radzishevsky A.Yu. Fundamentals of analog and digital sound. - M .: Publishing house "Williams", 2006 - p.105-109).

The output of block 4 is connected to the inputs of the block for calculating the intensity of sensation of the original speech signal in the critical bands of segment 6 and the block for calculating the thresholds of noise masking in the critical bands of the segment of the original speech signal 7. The outputs of blocks 5 and 6 are connected to the input of the block for calculating the difference between the original and distorted values of intensity sensations 8. The outputs of blocks 7 and 8 are connected to the input of the block for calculating the modified distortion of the spectrum of barks 9.

Units 5, 6, 7, and 8 are described in (W. Yang. Enhanced Modified Bark Spectral Distortion (EMBSD): An Objective Speech Quality Measure Based On Audible Distortion And Cognition Model. / A Dissertation of the Requirement for the Degree Doctor of Philosophy . - May, 1999, pp. 63-75).

The procedure for calculating MBSD in block 9 is described in (W. Yang, M. Benbouchta and R. Yantomo Performance of a modified bark spectral distortion measure as an objective speech quality measure, IEEE ICASSP, pp.541-544, Seattle, 1998).

The output of block 9 is connected to the input of the block for calculating the minimum modified distortion of the spectrum of barks 10.

The output of block 10 is connected to the input of the codec storage block 11.

Procedures that describe the functioning of blocks 10 and 11 are presented in (O.I.Shelukhin, N.F. Lukyantsev Digital processing and transmission of speech. - M .: Radio and Communication, 2000 - P.108-112).

The industrial applicability of the introduced elements is due to the presence of the element base on the basis of which they can be performed.

A device that implements the claimed method operates as follows. In block 1, the next segment of the synthesized speech signal is formed in the quasistationary section, then it goes to block 3 in which the values of the spectrum of barks in the critical bands of the segment of the synthesized speech signal are calculated, information about these values goes to the input of block 5 in which the sensation intensity is calculated synthesized speech signal in the critical bands of the segment. Information about the intensity of the sensation of the synthesized speech signal in the critical bands of the segment from block 5 is then sent to the first input of block 8. In block 2, the next segment of the original speech signal is formed in the quasistationary section, then it goes to block 4, in which the values of the spectrum of barks are calculated in the critical bands of the segment of the original speech signal, information about these values is fed to the inputs of blocks 6 and 7 in which the intensity of sensation and sometimes in noise masking in the critical bands of the segment of the original speech signal, respectively.

Information about the intensity of the sensation of the original signal in the critical bands of the segment from block 6 is then sent to the second input of block 8, in which the difference between the original and distorted values of the intensity of sensation is calculated. From block 8, information about the difference between the original and distorted values of the sensation intensity is supplied to the first input of block 9, and information on the noise masking thresholds in the critical bands of the segment of the original speech signal coming from block 7 is supplied to its second input. In block 9, the MBSD estimate is calculated , information about this assessment is submitted to block 10, in which the accumulation and calculation of the minimum of MBSD estimates during the iterative search is performed. From block 10, information about the iteration number corresponding to the minimum MBSD estimate for block 11, which is designed to store the best linear codec codec parameters within the given limits.

The advantages of the method include the fact that vocoders with linear prediction eliminate the mismatch between the human hearing system during the perception of synthesized speech and the proximity criteria used that determine the rules for analyzing the suitability of the selected codec parameters when implementing the analysis procedure through synthesis. Experimental tests were carried out according to (GOST R 51061-97. Systems of low-speed voice transmission through digital channels. Speech quality parameters and measurement methods. - M .: Gosstandart of Russia, 1997. - 230 p.), Which showed that the application of this method allows increase the subjective quality of perception of synthesized speech by an average of 0.11 points.

The above technical solutions show that a device embodying the invention in its implementation is capable of providing a better perception of synthesized speech in linear prediction vocoders, the implementation of which is based on the implementation of the analysis procedure through synthesis.

Claims (1)

A way to improve the perception of synthesized speech when implementing the analysis through synthesis procedure in linear prediction vocoders, based on the fact that the transmitting signal is iteratively synthesized on the transmitting side for the duration of the quasi-stationary part of the speech, changing the codec parameters at each iteration in accordance with the linear prediction algorithm used to calculate the best in the framework of the given restrictions of the parameters of the codec with linear prediction, characterized in that iterative calculation of the criterion m differential distortion of the spectrum of the barks, changing the codec parameters in accordance with the linear prediction algorithm used at each iteration, then the minimal perceptual distortion is calculated, and the codec parameters corresponding to this iteration are considered the best and are used to synthesize a speech signal over the duration of the speech quasistationary portion on the receiving side .

Images