Heteronyms of homographs = No

I feel like I click “No” so often, I thought I’d verify one type is indeed a “No.”

For a contrived example, I would hit “No” if I heard: sub‑MIT a new re‑ZOOM. Sure, the verb “to resume” is pronounced re‑ZOOM, but by context the word shown is REH‑zuh‑may: résumé, even though it is written, as it often is in English, with no diacritical marks.

(Sorry for the elementary-school-style pronunciation transcription, but IPA is hot garbage.)

My latest actual example read: “The wheelhouse projects to the west of the pump house and has weatherboard sides.” One has to read the sentence in advance to know that “projects” is a verb. This contributor pronounced the other word that’s spelled “projects” …as if there were some projects to the west.

“No”, yes?

I think the general problem is that a useful transcription program should forgive people who speak “incorrectly” but shouldn’t be trained on sentences that have been read incorrectly.

So unless you’re sure that people of the given accent wouldn’t actually speak that way organically, you should probably press the skip button rather than reject the sentence.

1 Like

It shouldn’t make a difference to the DeepSpeech acoustic model if a word is pronounced correctly but in the wrong context. However, there may be other dataset uses for which this is a problem.

Generally, in these situations I am less forgiving of native speakers and more forgiving of those with accents, as I can’t always say with certainty that it is not a valid pronunciation for that accent.