Speaker dependent speech enhancement using sinusoidal model

In many conventional speech enhancement methods, discrete Fourier transformation is used in analysis, modification, and synthesis stages without incorporating a signal-dependent model or the prior knowledge about the underlying speaker characteristics. In this work, we integrate a sinusoidal model as speech signal model and further include speaker information captured in a trained speaker model in the form of a sinusoidal coder. We design a postfilter as a post processor after a conventional speech enhancement stage. We show that the proposed method significantly improves the perceived quality in particular for non-stationary noise and low signal-to-noise ratio scenar-ios. The improved performance predicted by instrumental metrics is further justified by subjective listening tests.