Voiced - Unvoiced - Silence Classification via Hierarchical Dual Geometry Analysis

The need for a reliable discrimination among voiced, unvoiced and silence frames arises in a wide variety of speech processing applications. In this paper, we propose an unsupervised algorithm for voiced-unvoiced-silence classification based on a time-frequency representation of the measured signal, which is viewed as a data matrix. The proposed algorithm relies on a hierarchical dual geometry analysis of the data matrix, which exploits the strong coupling between time frames and frequency bins. By gradually learning the coupled geometry in two steps, the algorithm allows for the separation between speech and silent frames, and then between voiced and unvoiced frames. Experimental results demonstrate the improved performance compared to a competing algorithm.