Predictive coding of speech: Historical review and directions for future research

The reasons underlying the success of predictive coding are outlined. Proper attention to auditory perception and masking of the quantizing noise have resulted in coders with high quality at 0.5 bits/residual sample. Ever more powerful chips have transported these basic principles into the realm of real-time reality. For sustained progress, current research should focus on 1. The effects of dynamic masking: during and immediately following rapid transitions in the speech spectrum the ear seems more tolerant to rough coding. The saving in average information rate could be substantial. 2. Better algebraically structured codes: codes are needed that are fast, simple to implement and allow the incorporation of subjective criteria. 3. Efficient coding of the parameters: because of the great success with coding the residual, sparse representation of the predictor coefficients is now paramount.