5th International Conference on Spoken Language Processing
Speaker recognition is usually accomplished by building a set of models from speech of a known speaker, training data, and subsequently using a pattern matching algorithm to score the speech from an unknown speaker, test data. In this paper we discard the notion of train and test data in speaker recognition and introduce the multilateral scoring technique. This technique comprises building speaker models on material for the known speaker and matching the unknown speaker data to these models, the traditional approach to speaker recognition. The resultant scores are fused with an equivalent set of scores produced by matching the known speaker utterance to models built on the unknown speaker data. Significant improvements have been achieved using this technique on the NIST 1996, 1997 and 1998 Speaker Recognition Evaluation data. Results are presented for two speaker recognition systems, the first based on Hidden Markov models and the second based on Gaussian Mixture models.
Bibliographic reference. Parris, Eluned S. / Carey, Michael J. (1998): "Multilateral techniques for speaker recognition", In ICSLP-1998, paper 0444.