The Research Of Speech Digital Coding Technology And Effect Of Quantization Noise On Signal Reconstruction In The Compressed Sampling Enviroment

When facing the ultra-wideband signal or the signal with high level of redundancy, thetraditional signal processing framework based on the Nyquist sampling theorem not only results ina waste of sampling resources, but also deals inefficient. The theory of compressed sensing (CS)proposed in recent years. By utilizing the the sparsity of transmitted signal, CS technique canprovide signal compression and sampling simultaneously. As it brings the complexity fromsampling to reconstruction, CS technology greatly reduces the cost of the sampling signal, andbecome a hot research field of signal processing. However, for the digital communication system,sampling is only the first step for singnal digitalization, and the key step lies in the quantizationcoding for the sampling signal. Few of current research works involve the processing for themeasurements. With this background, this thesis investigates the speech coding problem in thecompressive sampling environment. Note that this work is the premise before the application of CStheory in real signal processing systems.First of all, leveraging the speech signal sparse pretreatment technology, this thesis proposed avector quantization coding scheme based on compressed sensing and the sparse pretreatmenttechnology. The main idea of the coding sheme is to code the overall sequences of themeasurements and reconstruct the original speech signal from the decoded measurements byutilizing the CS technique. Simulation results show that sparse pretreatment technology improvesthe speech sparse representation effectively, and compared with the vector quantization codingscheme based on compressed sensing without sparse pretreatment technology, this scheme can gethigher quality of synthetic speech under the same digital rate.Then, a speech coding scheme based on quantized compressed sensing is proposed. The mainidea of this theory is coding the measurements sequence by Lloyd-Max quantization andreconstructing the original speech signal from the quantized measurements sequence directly.Simulation results show that without decoding the quantized measurements sequence, this schemecan reconstruct the original speech signal from the quantized measurements sequence directly underthe premise of guaranteeing the satisfied quality of the reconstructed speech signal.To further reduce the amount of data, a speech coding scheme by employing matching pursuitbased on sinusoid modeling. We decompose the measurements sequence into few sinusoidmodeling bases with matching pursuit and output the parameters, which condense the information of measurements sequence. Simulation results show that the encoding method based on modelparameters can greatly reduce the data rate, but the synthesis speech quality remains to be furtherimproved.