Low bit rate speech compression using hidden Markov modeling

This paper presents a new approach for low bit-rate speech compression that exploits the acoustic structure of the speech signal to yield a reduction in bit-rate with negligible loss in speech quality. This new approach encodes the signal using subword units of decomposition that can be extracted through hidden Markov modeling. The subword decomposition yields a segmentation of the signal into variable length quasi-stationary segments that are encoded using a variable rate multimode compression scheme. In this paper, we show that quality is dependent on the type and structure of the subword unit chosen for decomposition and compression. We show that fenones, which correspond to acoustic subword units, provide better quality than linguistic-based units such as phonemes. The speech model is based on a harmonic representation similar to the sinusoidal transform coder developed by McAulay and Quatieri (1986, 1992) and the multiband excitation model developed by Griffin and Lim (1987, 1988)