This paper presents a low bit rate speech coder based on predictive lattice vector quantization (PLVQ) and time-scale modification (TSM). The coding model of proposed vocoder is built on the MELP, in which bit rate reduction is achieved by taking advantage of PLVQ and TSM techniques. PLVQ is used to encode the speech line spectrum pair (LSP) parameters, which has the advantage of lower implementation complexity than multi-stage vector quantization (MSVQ), moreover, it does not require memory for codebook storage. With our speech data base, PLVQ can save up to 4 bits/frame compared to unstructured codebook MSVQ. TSM can change the speed of speech signal with its perceptual characteristics remained. Through appending TSM as previous and post process, speech coding at bit rate about 1.1 kbps could be easily achieved without modifying the vocoder structure.
Weighted vector quantization of harmonic spectral magnitudes for very low-bit-rate speech coding