Reverse water-filling in predictive encoding of speech

Reverse water-filling suggests that, at low bit rates, the synthesis filter for predictive encoding should differ from the model filter of the signal to be encoded. However, reverse water-filling follows from optimum encoding and stationary Gaussian assumptions. By means of simple experiments, we show that reverse water-filling applies to predictive encoding of speech. For a vector analysis-by-synthesis encoding based on a first order autoregressive signal model, the use of a synthesis filter derived from reverse water-filling resulted in consistently improved segmental SNR measures.