Speech coding using nonstationary sinusoidal modelling and narrow-band basis functions

The authors present simulation results for several variants of speech modeling using sinusoids for voiced sounds and a representation by sinusoids or narrowband basis functions (NBBFs) for unvoiced regions. In most of the experiments the frequency range up to 4 kHz has been subdivided into 17 frequency groups (critical bands) equispaced according to a Bark scale. The goal of the simulations was to obtain some conclusions about suitable parameters extracted from the short-time spectrum of the input signal and their transmission for a speech coder working at medium-to-low bit rates. Best results were obtained from a relatively simple scheme using harmonically related sinusoids for voiced frames and straightforward NBBF modeling for unvoiced frames. No phase parameters were needed for synthesis. A novel pragmatic amplitude-correction approach called peak assignment yielded a clear quality improvement