A Multivariate Voicing Decision Rule Adapts to Noise, Distortion, and Spectral Shaping

This paper presents a new approach to making voiced/unvoiced decisions. The technique is very accurate and dynamically adapts to a wide variety of environments. Reliable decisions are achieved by using a decision rule to combine multiple speech parameters into a single discriminant variable. The algorithm is adaptive because it derives the decision rule from the incoming speech rather than from a training set. Multivariate clustering is employed to separate voiced from unvoiced speech segments and to determine voicing for each frame. An overview of related solutions and performance of the new method are presented.