The Research Of Speech Compression Coding And Reconstruction Technology Based On Compressed Sensing

We are in the information age, an interconnected world. The amount of information, the speedsof information processing and transmission grows rapidly. Classical digital system processes inputsignal based on Nyquist sampling, when facing broadband signal or signal with high level ofredundancy, it not only brings huge pressure to sampling system, but also results in a waste ofcomputing and memory resources. Therefore, Compressed Sensing came into the public view as anew data sampling method and became a hot research field of signal processing. Based on thetheory, this thesis studies the speech coding and reconstruction problem needed to be solved inpractical speech communication system.First of all, leveraging the speech linear prediction technique, this thesis constructed a PCMcoding scheme based on compressed sensing and linear predictive analysis. The steps includecoding on measurement sequences and linear prediction coefficients, reconstructing of linearprediction residuals with decoded measurements and linear prediction coefficients and recoveringthe original speech signal using residuals. Linear predictive technology improves the speech sparserepresentation effectively, thus reducing compression ratio of signal. Compared with traditionalspeech PCM coding, although the quality of synthetic speech is not improved, the digital rate is ableto be reduced.In order to further reduce the amount of data transfer, this thesis attempts to make a model ofmeasurements sequence and proposes a compressed sensing-based code excited linear predictioncoding scheme. Row echelon matrix is used to measure the original speech signal and I verify thatthe measurements sequence with row echelon measurements matrix is in accordance with the LPCmodel. This thesis then calculates the model parameters such as linear prediction coefficients, linespectrum frequency coefficients, pitch period, gain and proves the relationship between themeasurement pitch period and the original pitch period from a mathematical point of view. Finallythe overall measurements are encoded with Code Excited Linear Prediction encoding method.Experiment results show that sampling with half of the compression ratio, the encoding methodbased on model parameters can greatly reduce the data rate, but the synthetic speech quality remainsto be further improved.In the end, this thesis analyzes greedy algorithm’s reconstruction and anti-noise performancesin noise situation based on speech compression coding and reconstruction framework via wavelet decomposition. It proposes an improved adaptive compressive sampling matching pursuit algorithm(ACoSaMP).This algorithm can approximate the sparsity in different phases by setting the variablestep size in sparsity unknown circumstances. Meanwhile，in every iteration, the residual signalestimates are used instead of the least squares method for the whole signal, reducing algorithmcomplexity. Experiment results show the subjective and objective performances of the algorithmoutperform existing similar algorithms in different compression ratios and this algorithm has strongrobutness to noise.