Algorithms Of Speech Coding Based On Compressed Sensing

Compressive Sensing is a new theory that captures and represents compressible signals. It first employs non-adaptive linear projections that preserve the structure of the signal, and then the signal reconstruction is conducted using an optimization process from these projections.This paper introduces the theoretical framework of compressed sensing and discusses the existing difficult problems. Based on the analysis of the theoretics on compressed sensing and existing speech compressing coding methods, a new speech compressing coding algorithm using compressed sensing was presented. This algorithm is based on the characteristic of speech signal in frequency domain and the low frequency and high frequency coefficients were compressed respectively based on the information and energy they took. In addition, a new multiple description coding method for speech transmission was proposed. The method is robust to packet loss or bit error, and has the advantages of simple structure and easy implementation. The method made random measurements in the DCT or MDCT domain of the speech signal, and formed multiple descriptions after quantizing and packing. At the receiver end, the decoder can reconstruct the speech signal approximately or exactly with the received bit streams comparing with the single description algorithm by solving an optimization problem.The two designed algorithms above are fully tested and evaluated in compression ratio, the sparsity of the signal, quantization error and the influence to the algorithms using DCT coefficients or MDCT coefficients. All of the reconstructed speech signals are evaluated using the objective speech quality evaluation algorithm--PESQ. Experiments proved that the CS-Speech Compressing coding algorithm is able offer high compression ratio with small compression distortion and the CS-Multiple Description Coding can be used to transmit speech signals in real-time in unreliable networks, it is able to provide robust source coding algorithms without additional delay.