Scalable low bit rate CELP coder based on compressive sensing and vector quantization

Abstract

Code Excited Linear Prediction (CELP), one of the most famous hybrid speech coders, exploits the advantages of parametric coders and waveform coders. The quality of reconstructed speech increases with the size of Gaussian codebook used for quantizing excitation sequences. But this results in increased transmission bit rate and search complexity of the codebook. This can be dealt with using tools from Compressed Sensing (CS) domain that transfers complexity of transmitter to space of sparse recovery at receiver. Sparse signal recovery gained much interest in signal processing research as it allows data sampling below Nyquist rate. A Compressive Sensing based CELP coder that allows bit rate scalability by varying the dimension of the measurement vectors is designed and implemented in this paper. Vector quantization of CS measurements and Linear Predictive Coding (LPC) coefficients using Gaussian and LPC codebooks respectively resulted in a bit rate of 11.9kbps which is less than that of CELP coder of the same speech quality. By optimizing the number of bits allocated for parameters and interpolating the LP coefficients, the bit rate is further reduced to 8.1kbps without much degradation in the quality of the reconstructed speech.