Conversation
Ruff :(
Co-authored-by: michelwi <michelwi@users.noreply.github.com>
|
What a dangerous thing to do :-) I mean re-ordering the phonemes. But PR looks good to me. I am sure you tested it already. |
Indeed... However, this job has a purpose. If we don't use any state-tying file, the NN softmax output labels are obtained from the phonetic inventory. As a consequence, we need to have a consistent order between training and recognition lexicon phonemes so that we can output the same labels in recognition as the ones we learned in training. And I'm certain that I have a training/recognition lexicon pair for which consistent phonetic ordering is not happening. I hope we don't have to use this job much, but just in case it's beneficial for anyone or for some other use case, I decided to push it here.
Yes, I went through the phoneme set and it was ordered as expected! |
Asserts that the units in the phonetic inventory of the
ref_lexare equal to those of thelex_to_modify, and enforces the ordering in the phoneticlex_to_modifyto be the exact same ordering asref_lex.Useful if a given pipeline has use cases in which the order of the phonetic units inside the lexicon matters.