Joint Source Channel Speech Decoding using long-term residual redundancy

In this paper, we present a minimum mean squared error decoding scheme for reconstruction of encoded speech transmitted over a noisy channel, utilizing the residual redundancy due to the quasi-periodic property of voiced segments of speech. The proposed algorithm is in general applicable to any speech codec provided that the long-term memory is not exploited in the encoding process. An enabling step of the proposed solution is the pitch extraction based on the noisy source encoder output. This work is the first attempt in the concept of joint source channel decoding (JSCD) which uses the long-term residual redundancy of speech encoders output. Prior art has focused on the short term redundancy or the memory between successive symbols. Simulation results, based on the 32 kbps ADPCM codec of the ITU-T G.726 standard, indicate enhanced reconstruction performance when compared to the standard decoder.