Punnoose A K, Flare Speech Systems, India

This paper discuss an approach to detect whether a wave file contains speech or not. A frame classifier is trained to classify frames to phones. The inherent biases of the frame classifier, in terms of various aspects of recognition, is captured in terms of probability distributions. Using the distributions of speech and noise, an approach is presented, which scores wave file for the presence or absence of speech. Relevant databases are used to test the detection rate of this approach.

Noise Robustness, Neural Networks, Interactive Voice Response Systems, Confidence Scoring