Open artrayd opened 3 years ago
this is an issue caused by the loss function used in the model (CTC), unfortunately, it is known to have this peaky issue and cannot be fixed.
Hi @xinjli thank you for your answer. Another question, is that theoretically possible to mark pauses? A time when there is no voice at all?
I don't know what your use-case is @artrayd but we've had success combining the current audio volume with allosaurus output for generating animated lipsync. It patches over any sounds allosaurus does not recognise, and helps us respond to silence correctly.
However, if you use-case does not have clean enough audio for that, we have seen that allosaurus is remarkably good at just not outputting anything for periods with non-speech sounds. So simply finding gaps of a particular length in the allosaurus output may well be suitable.
@willstott101 thank you! I was thinking in the same direction.
I suspect that due to this issue, allosaurus "swallows" multiple phones when the speech is rather quick (in Estonian, for example, native speakers tend to produce sounds quickly because the words are so long). Might this be the case? If so, what is an CTC model, where do I learn more about it?
(I suspect that -e
option might be intended to compensate for that "fixed duration" thingy.. unfortunately it does not seem to do any better for the overall outcome)
No matter what, phone duration is 0.045 that doesn't sound right. Even if I say something like "Ooooooooh yeeeeees"
4.080 0.045 iː 4.320 0.045 tʲ 4.410 0.045 iː