"Confidence analysis" was added to a speaker identification procedure to enable us to get the confidence that the target is indeed present in or absent from the reference list. The degree of confidence was deduced from an analysis of the differences between two groups of curves, namely AvD (Accuracy versos Distance) and PAvD (Pseudo Accuracy versus Distance). AvD was calculated from the percentage of cumulative matches versus intraspeaker distance. PAvD was obtained the same way with the exception that a speaker's test vectors were not compared against his personal reference template. Instead, another candidate with the most pseudo matches was selected to calculate PAvD.
In this paper, we investigate the effectiveness of four sets of acoustic features (autocorrelation coefficients, reflection coefficients, cepstral coefficients, and log area ratio coefficients) applied to the confidence analysis. Experiments reveal that: (1) autocorrelation coefficients set is not effective, (2) reflection coefficients set generally excels against any other feature sets, and (3) for some specific speakers, log area ratio coefficients set and cepstral coefficients set demonstrate their respective superiorities.
Cite as: Ong, S., Moody, M.P., Sridharan, S. (1994) Confidence analysis for speaker identification: the effectiveness of various features. Proc. ESCA Workshop on Automatic Speaker Recognition, Identification and Verification, 91-94
@inproceedings{ong94_asriv,
author={Sherman Ong and Miles P. Moody and Sridha Sridharan},
title={{Confidence analysis for speaker identification: the effectiveness of various features}},
year=1994,
booktitle={Proc. ESCA Workshop on Automatic Speaker Recognition, Identification and Verification},
pages={91--94}
}