Recovery of speech spectral parameters using convex set projection

Previous works have demonstrated that by preserving speech spectral “dynamics” during spectral parameter quantization and/or decoding, the quality of coded speech can be improved. We explore the use of projections onto convex sets (POCS) techniques to recover speech spectral parameters from their quantized versions. Unlike prior works, the POCS approach enables us to obtain solutions that satisfy precise constraints. Two constraint sets are used in our POCS recovery algorithm: one set constrains the “roughness” of the parameter trajectories, and the other set confines the parameters to the proper quantizer partition cells. Simulation of our algorithm has consistently produced improvements in both the subjective quality and objective distortion measurements