Abbreviations in dataset transcript

[JRMeyer]

>>> jerry.matjila
[October 16, 2020, 7:52am]

1. When creating a dataset, what is the be best way to transcribe
abbreviated words? slash
Here is list of few examples: slash
CAN which sounds 'see a n' rather than 'can' - should it be 'C A N' slash
BEE which sounds 'be e e' rather than 'bee' - should it be 'B E E'

2. what is the be best way to transcribe a year? slash
Example 1985. Should it be nineteen eighty five or simply 1985?

Your advice would be highly appreciated.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/abbreviations-in-dataset-transcript]

[JRMeyer]

>>> nmstoker
[October 17, 2020, 12:30am]

With 2 it's safest to transcribe it to words as it was actually said,
since then you'll be sure that no matter how the text is processed it
will match. This is useful in cases where it is ambiguous, like 2012
where sometimes it's said 'two thousand and twelve' and others 'twenty
twelve'

With 1 it depends on how phonemizer would say it - you want the way
you've transcribed it to be turned into the same sounds/words by
phonemizer as your recording actually used. slash
I'm not by my computer to test it, but I think generally with English
abbreviations when using espeak-ng through phonemizer it would generally
turn capitalised words into the equivalent of the letters. Thus 'BBC'
becomes the equivalent of 'be be see' and provided that was how it was
said then it would be fine to transcribe as BBC in that case.

[Archived Post]