Robust voiced/unvoiced/mixed/silence classifier with maximum a posteriori channel/background adaptation

A new statistical voiced/unvoiced/mixed/silence classifier based on a maximum a posteriori (MAP) adaptation algorithm is presented. The speech signal distributions are modeled with Gaussian mixture models (GMM). The MAP re-estimation of model parameters is based on the sufficient statistics within a form of Bayesian adaptation. The robustness of the proposed technique and model adaptation to different background/channel conditions were evaluated. Experimental results show that the proposed method can adapt and is robust to adverse signal conditions, such as SNR as low as 3 dB, noise in a moving vehicle, and band-limited channel conditions.