Improvements of Memory Vector Quantization for Noisy Channel Transmission of Lsf Parameters

This paper addresses the problem of efficiently transmitting the LSF speech parameters over a noisy channel using vector quantization (VQ). A comparison of some traditional memory VQ methods is performed, where we investigate what gains can be achieved by exploiting interframe correlation. By combining a predictive VQ with a fixed memoryless VQ, called the safety-net, further improvements in performance are obtained. Some methods for reducing the effects of channel errors are also investigated. We show that many memory methods perform worse than memoryless VQ when the channel is noisy, but the SN-PVQ outperforms memoryless VQ for all tested channel error rates, with 4 bits less.