Revisiting the Linear Prediction Analysis-by-Synthesis Speech Coding Paradigm Using Real-Time Convex Optimization

In this work, we propose a novel approach to speech coding by rewriting the nonlinear analysis-by-synthesis linear prediction scheme as a convex problem. This allows for determining trade-offs between, on one hand, the reconstruction error and, on the other, the sparsity of the predictor and the residual used to parametrize the speech signal. Differently from traditional coding schemes where the parameters are chosen throughout multiple optimization stages, our scheme produces a one-shot parametrization of a speech segment that intrinsically takes into consideration the voiced or unvoiced nature of a speech segment providing a better balance between residual and predictor and, consequently, a more appropriate bit allocation.