Non-linear interpolation in prototype waveform interpolation (PWI) encoders

A lack of fidelity of periodicity over voiced regions is considered to be an important factor in the performance degradation of CELP class speech coders operating at or below 4 kb/s. Improvements in vocoder performance obtained by using extra bits to represent pitch waveform structure indicate that improved waveform fidelity is of equal importance. Prototype waveform interpolation (PWI) schemes provide an efficient tradeoff between these two requirements. PWI recovers voiced speech from only one pitch cycle waveform (the prototype) per frame. prototypes are transmitted to the decoder and the remaining voiced speech is reconstructed by interpolation from the prototypes. Although the interpolated speech differs from the original this approach results in surprisingly good quality at bit rates at or below 3 kb/s. In principle, only nonlinear procedures can succeed in forcing the interpolated signal to approach the target with unlimited precision. However, in low bit rate speech coding, the associated penalty in terms of additional bit requirements is crucially important. There are inherent problems in maintaining waveform fidelity using linear techniques and the work reported here is aimed at providing a better representation of periodic structure using nonlinear interpolation functions. Waveform fidelity within interpolated regions is taken into account more explicitly by employing a two-stage nonlinear interpolation process