Open ML6634 opened 3 years ago
Yes, you need to add into lexicon mapping {laugh} {laugh}
so that it will map these words into the token itself and not using anything else.
In list file you can have any text, the thing you need to work is lexicon, how you map words into tokens set.
To have sounds made by the talker, background or channel, for example, {laugh}, {cough}, {breath}, [background noise], [channel noise] etc. and those "words", in transcriptions on the list files, such as %mm, uh-huh, and %um etc. to be considered, I just need to append, for example,
{laugh} {cough} {breath} [background noise] [channel noise] %mm uh-huh %um
to the set: librispeech-train-all-unigram-10000.tokens. Right? After this, do I also need to modify lexicon and vocab or not?
In addition, in transcriptions on the list files, at the front of the name of a person or a place, there is an &, for example, &Jack, &Chicago etc. How do I take care of it? Just add
%
to the set of tokens? If so, when an audio is transcribed, may the model recognize a name and prefix it with &? Or, is it better if before the training I remove the prefix, &, from the names in transcriptions on the list files? Thank you!!