The effects of noise on the waveform interpolation speech coder

摘要

A study of the effects of additive noise on the WI coder is presented. In WI, pitch-cycle waveforms (prototypes) are extracted from the residual signal. Since these prototypes evolve slowly when the speech is voiced and rapidly when unvoiced, simple filtering can be applied in the evolution domain to separate these components. The focus of this work is to investigate how this decomposition maps noise, with application to speech enhancement. A repercussion of the alignment procedure, necessary to align successive prototypes, causes incorrect decomposition of narrowband noise, resulting in the spreading of energy. Furthermore, the near equal decomposition of broadband noise into the slowly evolving and rapidly evolving component poses difficulties for noise suppression. Thus, problems with the current WI coder in noisy conditions are identified, highlighting the necessity to develop further methods to overcome such problems.