It's my bad. WhisperX uses DTW for an accurate mapping of word level timestamps via wav2vec2. I have read somewhere (but can't find it) that using MFA instead performs much better, albeit that you need accurate transcription in the first place.. we'll trial various options and shout about it soon..

Thanks for this. Normally the original implementation has an MFA model attached to allow for forced alignment, non? In any case, we're working on reimplimenting it to work on non standard forms (especially Arabic dialects, which are non-written dialects). We'll trial it! Cheers

Léa Salamé "Une des choses que j’ai appris"
@arthurmensch (@mistral) "La technologie qu’on a mis à disposition "
L'accord du participe passé, ce n'est plus une variante de prestige, c'est juste une règle morte.
Et le non-accord permet de reconnaitre les humains !
@tract-linguistes.org

I agree with you. But it is that initial cost that can become prohibitive at some point!

But can't we combine null hypothesis testing with confidence intervals? I know bayesian is best, but still requires a lot of fine tuning. For instance, going for flat or non-informative priors as one of the easiest to apply, but requires proper assumptions in the data distribution, etc

But what's the alternative?

And here is a link to our paper describing the system www.lrec-conf.org/proceedings/...

You cam also use the WebMinni which provides a speech to text blind transcription it can also speed your work