Vector quantization of speech LSP parameters using trellis codes and l/sub 1/-norm constraints

Trellis-coded vector quantization (TCVQ) is used to encode line spectrum pair (LSP) parameters. Three intraframe encoding schemes are considered: direct encoding of the LSP vectors, encoding the differences of LSP parameters, and using nonlinear prediction in encoding the LSP vectors. The last encoding scheme is the best; it can achieve 1-dB spectral distortion using about 27 bits for each speech frame. An alternative LSP coder based on an l/sub 1/-norm constrained vector quantizer is also described. This two-stage VQ encoding scheme uses 25 bits/frame to achieve about 1-dB spectral distortion.