First International Conference on Spoken Language Processing (ICSLP 90)
It is recognised that some parts of speech contain more speaker discriminating information than others. The aim here is to automatically identify useful speech segments in order to improve speaker recognition performance. Initially we present a brief overview of our earlier papers on this subject, in which the speech classification idea is first introduced. We show the need for speaker-dependent classification and further examine the use of hierarchical classifiers to improve recognition performance. Finally a majority voting system is introduced, which uses person-specific classifiers in order to automatically reject utterances containing little speaker specific information. We show that this scheme can be used to reduce the identification error rate from 11.7% to 3.2% by the automatic rejection of 38.5% of the test utterances.
Bibliographic reference. Eatock, J. / Mason, J. S. (1990): "Automatically focusing on good discriminating speech segments in speaker recognition", In ICSLP-1990, 133-136.