The confidence scores may be interpreted as likelihood ratios.

The system is based on the well-known GMM/UBM paradigm.


For the training of the UBM model composed of 512 Gaussian components,
we used the following databases: Switchboard 2 Part 3, Swicthboard
Cellular 1, Swicthboard Cellular 2 and a subset of the NIST SRE
2004. For the features we used 16 LFCCs with their deltas and the
delta of the (log)energy (altogether 33 features), normalized to
conform to zero mean and unit variance distribution. The silence
removal was based on a simple GMM based energy detector.

The speaker (target) models are adapted (MAP adaptation) from the UBM.

The score for each trial was calculated as a ratio between likelihood
of the target and likelihood of the UBM model. No score normalization
was performed.

For the development we used the NIST SRE 2005 data.

If the processing had been performed on a single CPU (3.2 GHz P4) with
1GB of RAM, the approximate CPU time spent would be around 12 hours
for training of the target models and around 48 hours for processing
of the test data.