Robust transmission of MELP-compressed speech: a tale of two channels

This paper considers the transmission of speech compressed using the US federal standard 2400 bps compression algorithm. Two different kinds of communication channels are considered - a noisy channel characterized by fading and additive white Gaussian noise, and an ATM channel subject to cell loss. Two different error control strategies are applied to these two channels. For the noisy channel, standard convolutional codes are employed with a channel decoder "tuned" to exploit the residual redundancy inherent in the compressed bitstream; considerable coding gain is obtained with this approach compared to the standard decoder that does not exploit residual redundancy. For the lossy channel, Reed-Solomon codes are used with erasure decoding to recover lost cells; it is shown that a modest investment in interleaving and redundancy can yield near-noiseless performance even when the channel is subject to cell loss as high as 10-15%