Research On Speech Digital Coding Based On Compressed Sensing

The arrival of "big data"(Big Data) era has brought rapid growth of data and information processing tasks. The minimum sampling rate of conventional Nyquist sampling theorem not only makes the sampling equipment complex, but also causes a waste of system recourses in the subsequent storage and compression stage, when we process the ultra-wideband signal or redundancy signal. Compressed sensing(CS) technology can be implemented to do signal compression while the sampling process, which has attracted widespread attention in the field of signal processing in recent years. For a complete digital voice processing system, quantization coding is an important part of digitalization. However, the study of compressive sensing is still in the initial period, few research works like encoding have been involved in the processing of measurement sequences. As a substitution of Nyquist Sampling technology, signal coding is a necessary prerequisite for compressed sensing’s leap from theory to practical application. This thesis mainly works on the digital coding methods of speech measurements based on compressed sensing. The main works and innovations of the thesis are summarized as follows:(1) Among the classical Nyquist sampling based speech coding techniques, model-based coding has obtained wild applications. Inspired by the traditional speech sinusoidal model, the thesis firstly utilized a dictionary of sinusoidal atoms and orthogonal matching pursuit algorithm for compressed sensing’s measurement modeling. For amplitude, phase and frequency, three kind parameters of each frame, using appropriate encoding techniques according to signal sequence’s characteristics, which could improve transmission efficiency apparently. In the decoder, using the decoded parameters to recover measurement. Basis pursuit algorithm is utilized to reconstruct the synthesized speech signal, and a rear low-pass filter would improve human’s auditory effects of the synthesized speech. Simulation results show that the coding scheme achieves compression coding of speech measurement, and both subjective and objective quality of reconstruction signal could be guaranteed.(2) Take advantage of that the measurements can retain some characteristics of speech time domain’s features after projection by row echelon matrix, we employ sparse representation to model the measurements and design a new speech CS coder and decoder. Firstly in the training phase, adopting K-Singular Value Decomposition method and a large number of measurements to generate a speech measurement dictionary; then in the encoding phase, a small number of atoms are selected for representing real-time speech measurements, their locations and magnitudes are encoded and transmitted; decoder uses the quantized measurements and reconstruction algorithm to reconstruct original speech signal. The simulation results show that speech measurements coding scheme based on sparse representation can effectively reduce the transmission bit rate, while it can ensure an excellent quality of the reconstructed speech.