Temporally overlapping segment quantization for speech coding
Masaaki Honda · The Journal of the Acoustical Society of America · 1988
A method for the efficient segment quantization of LPC spectral parameters is presented for very low-bit-rate speech coding. Recent methods of segment quantization for speech coding have used a concatenation segment model, in which spectral parameters are represented as a concatenation of variable-length code spectral segments. The present method uses a temporally overlapping segment model to capture more efficiently the variance in the spectral parameters due to coarticulation effect in speech. Spectral parameters are represented as a sequence of temporally overlapping segments, in which each code segment is decomposed into fixed-length temporal patterns and their weighted patterns. A dynamic programming procedure is used to determine the optimum code segments and the segment positions of the overlapping model so as to minimize the spectral distance between the input and the model spectral parameters in the analysis window. An iterative algorithm for code segment design in this model is also presented. Comparison of the overlapping model with the concatenation model for speech coding [Shiraki and Honda, to appear in IEEE Trans. ASSP] is described in terms of spectral distortion and computational complexity.