Robust Voiced/unvoiced Classification Using Novel Features and Gaussian Mixture Model

Need for deciding whether a given frame of a speech waveform should be classified as voiced speech or unvoiced speech arises in many speech analysis systems. Several approaches have been described in the literature for making this decision. In this paper, we have presented two novel approaches of using acoustical features and pattern recognition. The first method is based on Mel frequency cepstral coefficient with Gaussian mixture model classifier, which resulted in approximately 90% identification accuracy and the other is based on LPC coefficient and reduced dimensional LPC residual with Gaussian mixture model classifier, which resulted in 92% identification accuracy. The performances of both approaches were compared for various levels of noise and optimum condition for training is determined.