![]() |
Dublin Computation Linguistics Research Seminars2006 - 2007 |
![]() |
|
Contact:
|
The Dublin Computational Linguistic Research Seminars (DCLRS) provides the opportunity for computational linguistic researchers to share their current work with Dublin-based colleagues and friends. Each year, one of the major third-level institutes of learning in Dublin:
Dublin Institute of Technology's (DIT) School of Computing has the honour as host for the 2006-2007 academic year. Next topic 27 April 2007 Dr. Marco Grimaldi Standard speaker identification (SI) approaches using some variant on Mel Frequency Ceptstral Coefficients (MFCCs), in conjunction with Gaussian Mixture Models (GMMs) or similar classifers, are not robust with respect to channel change. This can be helped in part by channel normalization, but this has other attendant problems. It has also been suggested that change in the speaking style, as in intentional disguise, or in a forensic context, may be likewise injurious to SI. We here present new results from the CHAINS corpus which provide examples of multiple styles and recording setups. We show that some style switches are not especially problematic for standard systems, as long as channel characteristics are held constant. We also show that standard approaches are able to deal with whispered speech. Whispering represents a limiting case, however, and we demonstrate that systems trained to identify speakers based on modally voiced speech do not generalize to whispered speech. Finally, we demonstrate that changing channel is considerably more harmful to SI procedures than most style changes.
The support for DCLRS 2006-7 given by the School of Computing, DIT, is gratefully acknowledged. |
|