“TalkPrinting”: Improving Speaker Recognition by Modeling Stylistic Features


Kajarekar, S., Sönmez, K., Ferrer, L., Gadde, V., Venkataraman, A., Shriberg, E., … & Bratt, H. (2003, June). “TalkPrinting”: Improving Speaker Recognition by Modeling Stylistic Features. In International Conference on Intelligence and Security Informatics (pp. 350-354). Springer, Berlin, Heidelberg.


Automatic speaker recognition is an important technology for intelligence gathering, law enforcement, and audio mining. Conventional speaker recognition systems, which are based on independent short-term spectral samples, suffer from a lack of noise robustness and are unable to model a speaker’s idiosyncratic stylistic features. This paper describes “TalkPrinting”, a program of research aimed at adding such stylistic features to conventional systems. Results on three preliminary systems based on stylistic features demonstrate that (1) the new features alone carry significant speaker information; (2) they also carry significant complementary information compared to the conventional features; and (3) they provide increasing improvements in performance with increasing test durations.

Read more from SRI