Analysis of Top-K Strategies for Open-Set Speaker Identification Applications
September 2012
David Colella, The MITRE Corporation
Fred Goodman, The MITRE Corporation
ABSTRACT
Recent performance gains in speaker verification systems suggest it is now viable to employ these systems in open-set
speaker identification applications where an automated decision is passed to a human-in-the-loop for final analysis and decision.
This paper examines the performance for when a speaker verification system expands into the identification domain. Our
results indicate that separate thresholds should be adopted for the verification and the speaker identification phases. Furthermore, adopting a "top-k" approach where the best k matches are passed to the analyst for final matching does not greatly improve system detection performance and has a significant impact on overall human workload.
Additional Search Keywords
speaker recognition systems, Bayesian interpretation, Bayesian analysis, Bayesian risk criteria
|
