The optimistic test was an alignment of the 240 dev-test male speakers. The MLP output probabilities were manipulated so that based on the alignment a probability of 1 was set for the correct phone and the other phone probabilities were set to zero. These probabilities were then used to rescore the HTK lattice. The result was a 33.3% WER (Speaker Stats).While we would expect there to be a 10% word error rate due to the limitations of teh lattice topology, this result may reflect the limitations of using the language scores from HTK.

The pessimistic test was essentially the opposite of the optimistic test. The MLP output was set for every frame to be the prior phone probabilities. The result of this test was a 73.5% WER (Speaker Stats). This shows the large degree to which all of our results are affected by the acoustic decoding done to create the lattices in the first place, and suggests caution in interpreting results in the vicinity of 70%. Essentially, 73.5% should be considered the baseline for our rescoring experiments.