This directory contains results for MixAR speaker recognition with modified nist database from the training evaluation part of NIST 2001.

Setup:

1) MixAR model with 8 components trained with full EM (not stochastic EM).

wrong_det_mixar_08.jpg

Result DET plot is in det_mixar_08.jpg. For comparison, Tang's baseline result with nist 2001 is also shown in red. No silence-locking component was used, or extra processing to use modified likelihood adjustments were done.

det_mixar_08.jpg

This one is after correcting the recipe file to read the correct input format - LITTLE_ENDIAN. Result DET plot is in det_mixar_08.jpg. No silence-locking component was used. A simple normalized log-likelihood (normLL) scoring is also used.  For a given MixAR model, the normalization is achieved by subtracting the LL of the training file corresponding to the model from the test file LL.