An extension to Fisher Linear Semi-Discriminant analysis for Speaker Diarization

Abstract

The Fisher Linear Semi-Discriminant Analysis is used in Speaker Diarization to project acoustic features into a discriminant and lower dimensional space. Given that such analysis uses short segments to estimate the scatter matrices, the projection could be improved by using longer segments (i.e., more information). Since a change of speaker is more likely to occur during periods of non-speech, we propose to use segments of speech produced by the boundaries estimated from a voice activity detection method based on Hidden Markov Models. Using datasets from the NIST Speaker Recognition Evaluations, we show that the estimated segments provide a better scatter matrices for the analysis. The results show a relative improvement of 21% in the Speaker Error Time on the Switchboard corpus used in the evaluations.