Is there a plan to release rejected audio for sentences at all?
I could have imagined it, but I’m sure I saw that idea discussed (although I can’t find it now )
This would be of interest for a few learning scenarios. My main one was to test some automatic sentence checking - I have “known good” examples but would be ideal to use some genuine “known bad” cases. I can create fake bad cases (eg corrupting the audio or swapping the audio with a different sentence text) but thought it good to ensure I had a representative sample.
When I get a bit further, I’ll put together a thread on the checking process, but the basic idea is to take one (or preferably a couple) of a “known good” sentence, run it through a syllable count of the audio and a few other things and then compare an unknown sample against that. The trick is to make it flexible enough to capture variations in how things are said, whilst still distinguishing audio that’s not correct. There are some Python tools based on Praat that seem helpful