Robust speech transmission over noisy channels employing non-linear block codes

In medium to low bit rate speech codecs the speech signal is represented by a set of parameters. The most important concept is presently code excited linear predictive (CELP) coding. A speech segment of typically 10 to 20 ms is described in terms of prediction coefficients, gain factors and excitation vectors. Due to the high compression rates (0.5-1.5 bits per speech sample) the parameters are partly highly sensitive against channel noise. In this paper we present a new error protection technique, that is based on a joint optimization of parameter quantization and a redundant non-linear block coding scheme. For parameter reconstruction, the principle of soft bit source decoding is applied. The proposed technique can be used in combination with conventional error protection such as convolutional coding and allows a flexible subdivision of the gross data rate for source coding and error protection