On tree codebook design for vector quantization of speech

Two tree codebooks designed following both a forward and a backward path are discussed and their performances are compared. The algorithm used is a systematic procedure that minimizes the guesswork during codebook initialization. Two different features, LSFS (line-spectrum frequencies) and reflection coefficients, are used in the design. The results showed that S/N (signal/noise) performances are close for the two tree codebooks and the full-search codebook. However, the codebook designed using LSFS has a significant gain in S/N over that designed using reflection coefficients