First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Time-Domain Flexible Matrix Quantization for Very-Low-Rate Speech Coding

Hirohisa Tasaki, Kunio Nakajima

Information Systems and Electronics Development Laboratory, Mitsubishi Electric Corporation, Kanagawa, Japan

The block quantizers such as a matrix quantizer and a segment quantizer are used in low-bit-rate speech coding systems. It has been demonstrated that block quantizers have good benefits in bit rate when the codeword label is transmitted with duration time information. Although, if we send speech information in fixed bit rate, the quantizer must have some fixed boundaries between quantized matrices, and this limitation induces a great deal of degradation on quantization performance. To solve this problem, we propose a new spectral quantization technique called the time-domain flexible matrix quantization (TFMQ) which based on the fact that listeners are insensitive to the degradation by a small amount of time-warping transformation. Subjective and objective quality evaluations on TFMQ have been performed for 40 short Japanese sentences talked by one male speaker. The results have shown that TFMQ has superoir perceptual performance than conventional matrix quantizers. And it have been shown that the improvement is mainly brougt by the time-warping transformation in the manner of minimizing the matching distortion, and the slight difference in the time-domain structure is absorbed efficiently by this transformation.

Full Paper

Bibliographic reference.  Tasaki, Hirohisa / Nakajima, Kunio (1990): "Time-domain flexible matrix quantization for very-low-rate speech coding", In ICSLP-1990, 109-112.