5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

A Comparison of Fusion Techniques in Mel-cepstral Based Speaker Identification

Stefan Slomka, Sridha Sridharan, Vinod Chandran

Speech Laboratory, Queensland University of Technology, Australia

Input level fusion and output level fusion methods are compared for fusing Mel-frequency Cepstral Coefficients with their corresponding delta coefficients. A 49 speaker subset of the King database is used under wideband and telephone conditions. The best input level fusion system is more computationally complex than the output level fusion system. Both input and output fusion systems were able to outperform the best purely MFCC based system for wideband data. For King telephone data, only the output level fusion based system was able to outperform the best purely MFCC based system. Further experiments using NIST'96 data under matched and mismatched conditions were also performed. Provided it was well tuned, we found that the output level fused system always outperformed the input level fused system under all experimental conditions.

Full Paper

Bibliographic reference.  Slomka, Stefan / Sridharan, Sridha / Chandran, Vinod (1998): "A comparison of fusion techniques in mel-cepstral based speaker identification", In ICSLP-1998, paper 0123.