Nonparametric rank-order statistics applied to robust voiced-unvoiced-silence classification

This paper describes a theoretical and experimental investigation for detecting the presence of speech in wide-band noise. A robust algorithm for making the voiced-unvoiced-silence decision is described. This algorithm is based on a nonparametric rank-order statistical signal-detection scheme that does not require a training set of data and maintains a constant false alarm rate for a broad class of noise inputs. Two rank-order decision procedures are investigated, the Kruskal-Wallis and the multiple use of the two-sample savage statistic. The performances of these detectors are evaluated and compared to that obtained from manually classifying twenty recorded utterances. In limited testing, the average probability of misclassification of voiced speech for the Savage case was less than 6, 13, 28, and 55 percent, corresponding to signal-to-noise ratios of 30, 20, 10, and 0 dB, respectively.