Multi-Band Coherence Features for Voiced-Voiceless-Silence Speech Classification

This paper presents an accurate voiced-voiceless-silence classifier which can be used in many speech applications (coding, recognition,...). Thanks to autoregressive speech modelization, discriminate features calculated from the coherence between the signal and its prediction residue are proposed. We used Gaussian mixture models (GMM) to fit features statistics and we applied Bayesian classifier to decide on each frame class. We illustrate the effectiveness of the approach in noiseless and noisy environments and we compare the Bayesian classifier to linear discriminant analysis classifier