Achieving high quality speech at low bit rates: Fundamental issues in psychophysics

Abstract

Excellent progress has been made in recent years in synthesizing high quality natural sounding speech at bit rates as low as 8 kb/s. Rapidly increasing demand for speech transmission over digital radio channels requires that the bit rates be decreased even further—to 4 kb/s or lower. But, one is not able to achieve high quality speech at these low bit rates. Why? Low bit rate speech coders exploit auditory masking to mask the quantizing noise in the coded signal. The current models of auditory masking are valid only for relatively stationary segments, such as vowels, in the speech signal. Much of the distortions we heard in low bit rate coders are introduced during transitional segments, resulting from lack of psychophysical data on perception of such distortions. A human listener is able to detect such distortions in the coded version without comparing the coded version with the original. Why is it so? How does our hearing system know that the coded signal is distorted? Such questions must be answered before progress can be made in achieving high quality speech at low bit rates.