Compressed sensing based scalable speech coders

Code-excited linear prediction (CELP) is one of the most commonly used approaches for speech compression. The CELP coder, in particular the ACELP coder, relies on extracting an appropriate excitation sequence that is passed through a long-term and short-term linear prediction filter to synthesize the speech. One limitation of the analysis-by-synthesis search methods employed in the ACELP coder is that the positions of the non-zero entries in the excitation sequence and their gains are limited. Increasing the richness of the excitation to improve the speech quality is not only accompanied by the usual increase in the bit rate but also by a significant increase in search complexity. We propose dealing with this scalability issue by using tools from the compressed sensing (CS) domain. An analysis by synthesis coder using a CS based excitation sequence framework is developed and the encoder is shown to scale gracefully with CS dimension and offers a mechanism for quality to bit rate trade off.