A bifeature voiced/unvoiced discrimination algorithm for speech signals in the presense of noise

In this paper, a bifeature algorithm for the voiced/unvoiced (V/UV) discrimination of noise degraded speech signals is presented. A measure of merit for periodicity, and a normalized low frequency energy ratio are encompassed as two speech features which are proposed to derive from the linear prediction (LP) residual of the pre-processed noisy speech. Based on the statistically determined signal-dependent initial-thresholds of the features, a logical and multi-check V/UV classification criterion is developed which is capable of significantly removing the artifacts in the transitional frames between voiced and unvoiced regions. Through simulation results, a superior performance of the proposed V/UV discrimination scheme for noisy speech has been illustrated compared to some of the existing methods.