Nonlinear processing of phase vocoded speech

Although adaptive coding is in widespread use, the availability of very large scale integrated digital signal processing chips makes filterbank analysis and synthesis of speech signals very economical. This ear modeling simplifies the application of masking properties of the ear. Experiments were conducted to determine the number of filterbank outputs needed to reconstruct speech signals using a logarithmic bandpass filterbank consisting of an analysis and a synthesis filterbank. With eight outputs, the original speech sentence is found to be very clear; addition of 32 more outputs restores the breathiness to the sentence. A surprising degree of intelligibility is retained even with one output. The bandwidth of the speech signals is limited. The use of these filterbanks for speech enhancement and modest bitrate transmission appears favorable.