Research On Low Complexity Compressed Speech Sensing Codec For VoIP

With the rapid development of modern science and technology,the progress of voice call technology makes people's communication become more and more convenient.Voice over Internet Protocol(Vo IP)is a technology that uses the network to package Voice signals for transmission.In recent years,thanks to the rapid development of Internet technology,Vo IP technology has been more and more widely used because of its advantages of lower cost and good scalability compared with traditional telephone.However,Vo IP uses the network to transmit data.Therefore,network congestion and other unstable factors may occur,resulting in voice packet loss and delay,resulting in poor voice call quality.Although traditional codec methods such as interlacing technology can recover lost packets,the cost is the loss of part of the bandwidth.In recent years,studies have shown that compressed sensing technology can accurately recover lost data based on part of the data,so this technology has a broad research prospects in the field of Vo IP.However,the classical compressed sensing decoding method still has the disadvantage of high decoding complexity,so we urgently need to develop a low complexity compressed sensing decoding method to ensure the voice quality of Vo IP calls with low latency.The main research contents of this thesis are as follows:(1)The sparse characteristics of speech signals under different sparse representation methods are analyzed,and a dictionary learning method is proposed to further improve the sparse representation ability of speech signals.In the iterative process of solving the sparse coefficients,a certain amount of computation can be reduced by selecting multiple supports at one time,and a faster sparse coding method can be obtained.In the dictionary updating stage,each column of dictionary atoms is updated by the coefficients obtained in the fixed sparse coding stage.The experiment proves that the dictionary obtained by the dictionary learning method in this thesis has certain advantages in signal sparse representation ability compared with the traditional fixed dictionary,thus achieving better reconstruction effect.(2)The traditional method of speech codec tends to have high complexity,and the compression method has features of simple encoding of perception,the compressed sensing method is applied to speech decoding method,the use of refactoring faster In-crowd algorithm as a voice communication in the decoding side,which can get a lower complexity of decoding method and more conducive to voice real-time communication.In order to further meet the requirements of real-time communication for voice quality and transmission speed,a series of improvements are made to the In-Crowd algorithm: First of all,according to the third chapter analyses under the speech signal in wavelet signal will be most of the energy is concentrated in low frequency part,so take advantage of this characteristic,the low frequency part of the information as a priori knowledge on the In-crowd algorithm was improved,no need to look for support position of the low frequency part,part to reduce amount of calculation,and the reconstruction precision is improved.At the same time,due to the high computational complexity of the traditional basis tracking method to solve the convex optimization problem,the iterative soft threshold algorithm is used in this thesis to solve the optimization solution of the In-crowd algorithm,which can further improve the reconstruction speed.Through a series of experiments show that the proposed speech codec framework used in improving the low complexity of decoding method can achieve fast signal reconstruction,and in the use of this dictionary learning method to get the dictionary combines compression perception classic algorithms,although there is no advantages in the reconstruction speed,it can get a better voice quality.In reality,different network environments require different priorities for reconstruction speed and reconstruction accuracy.The speech codec method proposed in this thesis has certain advantages and feasibility in practical application scenarios.