Phonetic and Lexical Speaker Recognition in Reduced Training Scenarios
Brendan Baker, Robert Vogt, Subramanian Sridharan · 2004
High-level features have been shown to be effective for speaker recognition when large amounts of training data are available for speaker model training; however the feasibility of such long lengths of training for many applications is questionable. This paper describes the evaluation of phonetic and lexical n-gram based speaker recognition systems for reduced training lengths. Maximum likelihood modelling is compared to a recently developed MAP adaptation modelling technique. Results obtained using a restructured NIST 2003 Speaker Recognition Extended Data Task corpora indicate that significant gains in performance for both the phonetic and lexical based speaker recognition can be achieved through use of this adaptive modelling technique. The results from fusion experiments also demonstrated that the individual system improvements obtained for the high-level features translated into an overall performance gain when used along side traditional acoustic techniques. The MAP adapted modelling process was shown to extend the usefulness of high-level features to shorter training lengths, with results indicating that even when only one conversation side was used for training, the high-level systems provide complementary classifications and improved recognition performance. 1.