Non-linear prototype waveform interpolation for voiced speech encoding

Abstract

Prototype waveform interpolation (PWI) is a practical and promising coding technique applicable to voiced speech. The waveform and duration of only one pitch cycle (the prototype) per frame is extracted and coded using LPC techniques. Segments of missing speech between the prototypes are reconstructed at the receiver by interpolation from the decoded prototype waveforms. Although waveform reconstruction may not be very accurate over the interpolated segments, suprisingly good speech quality can be achieved at bit rates in the region of 2.5 to 3.5 kb/s using frames of about 20 ms duration, provided the prototype waveforms and pitch periodicity are satisfactorily reproduced. For reasons of low complexity and bit rate, most reported work on PWI uses linear interpolation methods with linear interpolation functions, but these suffer from inherent difficulties in reproducing non-uniform variations in pitch cycle waveforms and lengths. It is shown that nonlinear techniques can improve the representation of voiced speech in interpolated segments without significantly increasing bit rates. Pitch structure is improved by using a temporal differential rate codebook for transmission of small differences in the duration of pitch cycles. Waveform fidelity is improved by deriving optimal combination coefficients (OCC) which determine the composition of each pitch cycle waveform in terms of the given prototypes at segment boundaries. The OCC vectors allow for nonlinear variation in waveform composition and are vector quantised for transmission.