This directory contains results for MixAR speaker recognition with sustained phone database. (For each of the seven speakers, the database contains 2.9s from each of the 8 phones, all combined into one file for training and, 0.9s from each phone, all combined into one test file.)

Setup:

MixAR model with 8 components trained with full EM. No silence-locking component was used, or extra processing to use modified likelihood adjustments were done.

Result plots:

det1.jpg : DET plot
conv_EM_sustained.jpg : Convergence results for all the seven speakers.