Markov-switching Garchmodel and Application to Speech Enhancement in Subbands

In this paper, we introduce a Markov-switching generalized autoregressive conditional heteroscedasticity (GARCH) model in the short-time Fourier transform (STFT) domain. A GARCH model is utilized with Markov switching regimes, where the parameters are assumed to be frequency variant. The model parameters are evaluated in each frequency subband and a special state (regime) is de ned for the case where speech coef cients are absent or below a threshold level. The problem of speech enhancement under speech presence uncertainty is addressed and it is shown a soft voice activity detector may be inherently incorporated within the algorithm. Experimental results demonstrate the potential of our proposed model to improve noise reduction while retaining weak components of the speech signal.