INTERSPEECH 2004 - ICSLP
8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Triphone-Based Confidence System for Speaker Identification

Aaron Lawson (1), Mark Huggins (2)

(1) Research Associates for Defense Conversion, USA
(2) Lockheed Martin IT, USA

The confidence system proposed evaluates the reliability of Speaker Identification (SID) result based on the compatibility of the data involved in making the judgment. The measure employed in determining data compatibility is triphone overlap between the audio clip being tested, and the audio that served as training data for the model involved in the comparison. This approach evolved out of investigations into the degree of text-dependency of data used in human speaker verification. Those tests using text-dependant SID were more successful than non-text-dependant, presumably because subtle differences in a speaker's articulation were more evident and noticeable when similar or identical sound sequences were being compared. These findings were applied to automatic SID, using a measure of degree of text-dependency between training data for the model in question and audio clip being tested. The degree of overlap measure turns out to be a very accurate predictor of the SID success.

Full Paper

Bibliographic reference.  Lawson, Aaron / Huggins, Mark (2004): "Triphone-based confidence system for speaker identification", In INTERSPEECH-2004, 1745-1748.