A specially-collected speech database to reflect the long-term speaker variability.
F-ratio to determine the importance of speaker- and session-specific information.
Frequency warping and filter-bank outputs weighting strategies for feature extraction.