Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

Detecting User Speech in Barge-In over Prompts Using Speaker Identification Methods

Abraham Ittycheriah (1), Richard J. Mammone (2)

(1) IBM T. J. Watson Research Center, Yorktown Heights, NY, USA
(2) CAIP Center, Rutgers University, Piscataway, NJ, USA

In this paper, we investigate the use of a speaker identification technique to solve the bargein speechdetection problem. This scenario is a very simpleapplication of speaker identification since only twousers are involved. This is further simplified by thefact that the prompt speaker can be modelled apriori. Additionally, the user can be modelled as wellimproving the performance of the system on subsequent utterances. In the system described below, weexplicitly model several non-speech sounds such aslaughter, coughs and breath noises. We show thatthis technique is generally better than that of currentmethods which measure the ratio of incoming speechenergy to that of the prompt signal being played.


Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Ittycheriah, Abraham / Mammone, Richard J. (1999): "Detecting user speech in barge-in over prompts using speaker identification methods", In EUROSPEECH'99, 327-330.