HTK: Understanding scores in resulting .mlf file

Question

I am trying to understand the file result recout.mlf, so I have the following lines in that file:

Which of 'as' was well prononced: the one with -524.427185 or -1054.774536

Answer 1

The acoustic scores obtained during decoding are usually very tiny. To prevent underflow, log likelihoods are used instead of likelihoods: 1.5 Recognition and Viterbi Decoding.

Smaller argument values correspond to larger negative values of logarithms:

Thus, the first 'as' obtained a higher (-524.427185) acoustic score. Logarithm is a monotonic function (the larger is argument - the larger is the value), so you can compare the log-likelihoods directly: -524 > -1054.

BTW, it does not necessarily mean the first 'as' was better pronounced. The acoustic score depends on many factors, including model topology and the data the model was trained on.