Robust Adaptive Quantization: a Kalman-filter-based Approach

Adaptive quantizers/dequantizers are systems that are used to quantize efficiently signals with a large variation in short-term variance. They are typically found in telecommunication systems where highly non-stationary signals such as speech need to be represented digitally with the minimum number of bits. Channel errors that are introduced owing to non-ideal transmission significantly reduce the performance of adaptive dequantizers. In this paper we extend recent Kalman-filter-based adaptive quantization techniques to arrive at new dequantization schemes which are more robust to channel errors. This is achieved by utilizing the estimates produced by a Kalman filter based on a linear signal model which embodies the entire encoderlchannel combination. Extensions to Kalman smoothing based on the same signal model result in further performance improvement at the expense of a small delay.