Modified Viterbi algorithm for predictive TCQ

Summary form only given. A hybrid trellis-tree search algorithm, the H-PTCQ, which has the same storage requirement as PTCQ and, is presented. We assume 2 survivor paths are kept at each state. It is straightforward to extend the algorithm to the cases where n⩾2. Simulation is conducted over 20-second speech samples using DPCM, PTCQ and H-PTCQ. The data sequence is truncated into blocks of 1024 samples. The optimal codebooks for a memoryless Laplacian source are used. Predictor coefficients for the 1st-order and 2nd-order predictors are {0.8456} and {1.3435, -0.5888}, respectively. Simulation results indicate that both PTCQ and H-PTCQ have about 3 dB gain over DPCM. H-PTCQ with 8-state convolutional code has about 0.2 to 0.3 db gain over PTCQ for the same trellis size; H-PTCQ with 256-state convolutional code has 0.05 to 0.1 dB gain over the PTCQ counterpart. Compared with a 2M-state PTCQ, the M-state H-PTCQ has the same computational complexity and uses half of the path memory. Since the performance improvement of an an 8-state PTCQ over a 4-state PTCQ is about 0.4 dB for a similar set of data, the 0.2 to 0.3 dB gain obtained by using H-PTCQ is quite remarkable. Notice that H-PTQ enables a transmitter to adapt performance according to the resource constraints without changing PTCQ receivers. It is also interesting to observe that the 0.1 dB gain of an 8-state TCQ against a 4-state TCQ plus the 0.3 dB gain of H-PTCQ is about the gain of an 8-state PTCQ over a 4-state PTCQ. The results for 256-state quantization also agree with this observation. Therefore, we conclude that most of the gain of a 2M- over M-state PTCQ comes from the better internal TCQ quantizer, and mostly from the better prediction by keeping more paths