Ofer Hadar

Preceding research used the generalized likelihood ratio test (GLRT) for a robust voiced-unvoiced decision and within pitch tracking algorithm. The algorithm performance results were obtained using manual threshold evaluation. The purpose of this research is to propose automatic threshold evaluation methods. Constant false alarm rate (CFAR) and Bayes criterion thresholds are adapted for GLRT voiced-unvoiced decision and pitch tracking in the presence of white Gaussian noise WGN. The Bayes criterion threshold was also specially modified for babble noise conditions. An ad hoc method is proposed for threshold evaluation with no a priori probabilities required. The proposed thresholds performance is evaluated on voice signals contaminated by environmental noise. The results show that the proposed automatic GLRT threshold evaluation methods are robust to speaker and environmental conditions. The comparison between optimal manually chosen thresholds and proposed automatic methods demonstrates satisfactory error performance. The proposed Bayesian and ad hoc methods show excellent robustness and performance in the presence of babble noise compared to all other methods.