A new family of very low bitrate speech coders employing models of human perception is presented. The coding methodology is based on non-linear modulation of a random broadband noise source with signals derived from speech, following two main strategies for representing coded speech: one using the minimum audible difference between the original and the modulated speech signals, and another using a minimum log-error criterion along with some perceptually derived harmonic information. Depending on the methodology and the degree of accuracy employed for coding several implementations are allowed starting from 1 kb/s. High intelligibility is achieved even for the lower bitrate implementations, although, some increase in the bitrate is required for high-quality speech.
Bibliographic reference. Tsoukalas, D. / Mouropoulos, Jiannis / Kokkinakis, George (1995): "Very low-bitrate speech coding using perceptually-derived spectral data", In EUROSPEECH-1995, 233-236.