This directory contains models and statistics for evaluations. It includes a voice
classifier trained only on timestep 0 (noiseless samples). It was trained on the 360 hour
LibriSpeech dataset for 108K iterations with batch size 4 and lr 1e-4. This is roughly 1.3
epochs, so overfitting is unlikely to be severe. The final train log loss was 0.37 nats.
The model produces 512-dimensional features, and predicts 920 classes.

Also included is evaluation statistics, generated by stat_generate.py, for 10k training
set samples. The class score of the training set is 551.5.