Adaptivity Versus Tree Searching in DPCM

The interaction is studied between the effects of tree searching and predictor and quantizer adaptation in a differential pulse-code modulation (DPCM) speech digitzer. A 2 s male sentance is digitized at 1 and 2 bits/sample. The maximum entropy method computes a new source model every 8 ms, and a variety of strategiesare used to adapt the quantizer. Results are displayed m plots of local SNR versus time. It is found that the predictor, the quantizer, and the tree search each contribute roughly independently to the total SNR. Adaptivity contributes in total 6-7 dB and tree searching 2-3 dB. Local gains of both types can be higher.