The Selection of MELP Parameters to be Utilized with Joint Source-Channel Decoding

In this paper, we determine the necessary parameters of the mixed-excitation linear predictive (MELP) speech coder to be modeled with a first-order Markov chain in a joint source-channel decoding scheme. The selection is based on the bit error sensitivity analysis of the MELP bits and the calculated residual redundancy of the parameters. It is shown that although some parameter might not be essential for the speech quality intrinsically, the substantial residual redundancy of the same parameter utilized in the channel decoding can be beneficial for the overall system performance. We also measure the effect of mismatched training sequence to the joint-decoding of the MELP speech. The results show that even incorrect training sequence of different language provides considerable improvement to speech quality