Research On Modeling Technique Of Speech Signal Based On Compressed Sensing

With the growing ability of the sensor system to obtain data, people are increasing demand for information. Signal processing framework based on the required sampling rate and processing speed requirements are also raised. To address the problem of dramatic increasing pressure in the signal sampling process, people are working on a new sensing sampling model, compressed sensing theory, which is different from the traditional data sampling. In this thesis, we researched on the following topics based on the traditional compressed sensing:First of all, we analyzed the time series features , LPC model applicability and the non-linear characteristics of the observation sequence following the traditional compressed sensing theory. Then we concluded that the observation sequence is a kind of non-linear and compressed time series.Secondly, if the inputs of the signal processing systems are all sampled based on the compressed sensing theory instead of the traditional Nyquist sampling theory in the future, the modeling techniques for the observation sequence of compressed sensing will be a very important signal processing theory. The voice observation sequence is a kind of non-linear time series, so in this thesis we used several classical methods for non-linear time series modeling. At first, we applied the most widely used BP model to predict the observation sequence according to the advantages of the BP model and presented the compressed sensing theoretical framework based on BP neural network prediction model. Then we analyzed the prediction accuracy of the observation sequence and the final reconstruction accuracy of the speech signal based on the BP model through simulation.In the second part of this thesis, we analyzed the influence of quantization noise on the BP neural network prediction accuracy. That is because when we introduce the compressed sensing theory into practical digital communication system, it will inevitably generate quantization noises after the quantization coding process. Also, We finished three simulations to analyze the following topics respectively, which lay a foundation for the practical application of the compressed sensing: 1. the influence of the quantization noise on the traditional compressed sensing, 2. the variation trend of the SNR of the reconstructed signal with the quantization noise under the following two conditions: only inputs of the compressed sensing which is based on BP neural network model are distorted ;inputs and model parameters are both distorted.