Single channel speech dereverberation based on modified cepstral co-efficients

Linear prediction is the widely used method for obtaining all-pole speech representations. But in many cases spectral zeroes also plays a vital role in speech enhancement. So there is need of designing general modelling method for this purpose. Besides this, the requirement of pitch synchronization made the flexibility of existing algorithms limited. This paper exploits a novel approach on pole-zero analysis called cepstrum analysis, which can avoid synchronization problems. Dereverberation is performed based on cepstral separation on non-minimum phase signals. The technique works by flattening the magnitude of non-minimum phase signals prior to phase distortion. The proposed method is superior to standard linear prediction techniques as it overcomes the computational errors inflicted due to long time responses of all-pole inverse filters. The algorithm explores speech dereverberation techniques, whose principle is recovery of envelope variations of original speech (Langhans and Strube, Speech enhancement by nonlinear multiband envelope filtering, 1982; Hirsch, Automatic speech recognition in rooms, 1988). Results of the preliminary evaluation on reverberant speech using signal to reverberation ratio are analysed and discussed.