Mitigating the effects of channel noise in source compression with reduced rank processing

A technique is presented to reduce the effects of channel noise for source compression methods based on residual error linear prediction. The technique adaptively optimizes the prediction filter to minimize the reconstruction error in the presence of channel noise. The optimization is done using a form of reduced rank processing based on the multi-stage Wiener filter to control the impulse response of the prediction filter. By introducing estimates of the channel noise into the prediction filter design the overall reconstruction error can be minimized. The results show a 12 dB improvement on the reconstructed speech in noisy channel environments.