A wideband speech codec based on nonlinear approximation
M.G. Kokes, Jerry D. Gibson, Gerald Schuller · 2001
In transform-based wideband speech coding, the modulated lapped biorthogonal transform (MLBT) can be used to improve the frequency selectivity of the synthesis basis functions. In this work Campbell's coefficient rate and spectral entropy are used as a guide to develop band combining strategies to produce an adaptive nonuniform modulated lapped biorthogonal transform (ANMLBT). Due to the nonuniform nature of the transform, psychoacoustic quantization noise shaping is accomplished by employing time-varying pre/post filters. This new codec consists of a nonlinear approximation method to determine the best N basis functions that represent the current speech or audio frame. Preliminary coding results at an average rate of 16 Kbit/sec are demonstrated.