Convert arabic numerals and symbols to phonetic form

[m-bain]

Currently arabic numerals and symbols in whisper transcript cannot be aligned, needs to be phonetic alphabet.

Need to perform inverse of normalization in https://github.com/m-bain/whisperX/blob/main/whisperx/normalizers/english.py

Such that numbers and currencies are converted to their phonetic word form.

E.g.
"$300" -> "three hundred dollars"

To perform wav2vec alignment.

Then convert back to symbol form, and assign timestamps.

[m-bain]

update: finding this is a fairly ill-posed problem...

Given a numeric word from whisper transcript e.g.,

"2017"

I am finding this can be a many-to-one phoneme mapping:

"twenty seventeen", "two thousand and seventeen", "two zero one seven".

And so aligning these becomes a dense search problem. Then adding in currencies etc. (e.g. $2017), becomes even harder, too many possible things to align.

Alternative solution is to grab timetstamp from whisper attention score. Overall not sure if this is worth the time/complexity (since most words are unalignable and are not numeric)

[Ntweat]

Hi,

Any progress on this. ??

I tried the solution from "openai/whisper#1041" But I am still getting digit and not able to align them.

Anyway we can align the digits?