Open nefastosaturo opened 3 years ago
We could use library N2W-IT
I not sure where it is more convenient to implement this conversion: whether on the MITADS dataset (linguistic model) or in the MITADS-Speech trascription (acoustic model), or both. You think?
In MITADS, now, sentences that contain numbers are excluded (there are Roman number instead), they are excluded in merge_txt.sh Even the importers of MITADS-Speech, or in CommonVoice importer, these types of sentences are excluded.
looking better I realized that speech datasets we import, cardinal/ordinal numbers are not present, even if they are then processed in importers (e.g. importer CV). So it is correct to implement conversion in MITADS preprocessing text
We need to convert:
10 -> dieci 2012 -> duemiladodici 3,14 -> tre virgola quattordici
and so on
https://discourse.mozilla.org/t/converting-numbers-in-textual-form-to-numerical-values-in-stt-output/26321
then we can consider also: #100