Low Rate Encoding of Autoregressive Sources

Abstract

<p>Various approaches, traditional as well as non-traditional, are utilized to encode Gaussian and Laplacian distributed general autoregressive sources at rates of 1 and 2 bit per source letter. The performance of the traditional DPCM encoder is evaluated. At these low rates, DPCM turn out to be rather ineffective from a data compression point of view. Underlying laws governing the performance loss caused by the quantiser non-linearity in the predictor loop are detected experimentally. It is found that tree searching improves the performance substantially and the gain is a very well behaved function of some well known source statistics. Effect of tree searching on mismatched source predictor is examined: The results indicate that tree searching is not a substitute for a matched predictor. The performance of an intution-based smoothing filter in cascade with the DPCM encoder is evaluated when the predictor is matched as well as when mismatched to the source. Such smoothing is not helpful. Finally, a certain random coding scheme is used to rate 1. The performance of such as information theoretic inspired scheme is compared with the tree searched DPCM. Wherever appropriate, the relevance of results to low rate waveform encoding of speech is stressed.</p>