High Performance On-line Session-adaptation for Handling Inter-session Speaker Variability in Variable-text Speaker-recognition

摘要

We propose an on-line session-adaptation algorithm for making variable-text speaker-recognition systems robust to inter-session variability, by a continuous update of registered speakers' multiple templates from test utterances during actual usage of the system. The algorithm is set in a speaker-verification mode of operation and uses an enhanced verification step to ensure reliable selection of input utterances for template updation in an unsupervised way.