Exploring Intrinsic Speaker Qualities Via an Analysis of Automatic Speaker Recognition Systems
*** THIS PROJECT IS NO LONGER ACTIVE ***
Lara Stoll and Nelson Morgan
Although a great deal of progress has been made in the task of automatic speaker recognition, there are still many challenges that remain. A relatively limited amount of work has been done to investigate why system performance is better for some speakers than for others. We perform an analysis of automatic speaker recognition systems, with a focus on determining the inherent speaker characteristics that contribute to how easy (or difficult) it is to recognize a speaker correctly. We consider a range of intrinsic speaker qualities, including physical attributes, prosodic characteristics, and accents or dialects. Recent work successfully found impostor speaker pairs that are difficult for automatic speaker recognition systems to distinguish . The results of such analysis can then be used to improve speaker recognition systems. It may also be possible to predict the system's performance for an unknown speaker using only information about speaker characteristics. Currently we are developing an approach to do this for an open-set speaker identification task on noisy data.
- L. Stoll and G. Doddington, "Hunting for Wolves in Speaker Recognition," in Proc. of the Odyssey Speaker and Language Recognition Workshop, 2010.