Parametric segmentation of speech into voiced-unvoiced-silence intervals

In this paper a parametric technique for speech segmentation into voiced, unvoiced and silence intervals is described. This technique is based on the statistical properties of parameters which define a four dimensional pattern space. This space is spanned by the magnitude and the zero-crossing rate of the speech waveform and the normalized magnitude and the zero-crossing rate of the speech difference signal. A method is presented for the evaluation of the discriminating power of each parameter and the selection of the features for the decision criterion. The partition of the feature space is performed by discriminant functions parallel to the axes with a probability of misclassification less than 5%. The performance of this pattern classifier is evaluated and compared to that obtained by a reference criterion essentially based on the properties of the autocorrelation function.