Compressive sensing based scalable speech coder with dynamic selection of basis and vector quantization

Improving the sparsity of residue enhances the performance of Algebraic Code Excited Linear Prediction (ACELP) and many speech coders based on Linear Prediction (LP) due to the representation of excitation sequence. Compressive Sensing (CS) based signal recovery has witnessed an increased demand in signal processing research due to its low complexity at transmitter and novel sampling paradigm that does efficient data sampling much below Nyquist rate. Speech is a non-stationary signal whose characteristics vary greatly with time. A CS based coder that does dynamic selection of basis using voiced/unvoiced classification of frames is designed and implemented in this paper. This reduces the bit rate and encoder complexity by eliminating the need for evaluation and transmission of Linear Predictive Coding (LPC) coefficients for every frame. Variation in the dimension of random measurements introduces bit rate scalability into the coder. Performance efficiency of the coder is enhanced by applying vector quantization to LPC coefficients and random measurements by which the bit rate has been narrow down to 5.5 kbps. Further improvement in bit rate reduction is attained by optimizing the bit allocation to transmitted parameters that resulted in a lowered bit rate of 4.3 kbps with an affordable deterioration in the quality of reconstructed speech.