I got bit by this today with a podcast. Going forward, I'm only going to manually remove filler words.

Yep. Use descript for content edits and hire a professional for proper distraction editing. the AI will never know the right move with filler removal and how to sound natural

First time using Descript and I learned the hard way to (a) correct my transcript before trying to remove filler words because they might be misidentified and (b) don't bother using Descript to remove filler words because the results are very hit or miss compared to editing manually with, say, iZotope RX. It would be nice to be able to round-trip edits, perhaps by setting up markers and then re-importing audio, though I suppose the resulting audio I would import would be a manually "flattened" version of the original...

Totally agree. You've created an amazing product, guys! And many many thanks for your talent and passion... but yeah sometimes it works a little bit too aggressive, and one should double check its work.

Yes! And I'd like to add two related proposals that might help to mitigate this problem: (1) let users set a minimum confidence value for detecting filler words, and (2) let users set attack and release times for chopped-out filler words.

For (1), what I mean is that whatever machine learning model is being used to detect filler words probably labels each word with its level of confidence about that word being a filler word; and then Descript removes or ignores any words that have a confidence value higher than some threshold. It'd be nice if users could set that threshold value. In other words, I'd like to be able to specify something like: "Dear Descript, please only remove filler words if you're REALLY SUPER SURE that they're filler words." Sometimes, the model is a little too aggressive at detecting filler words, and I'd rather err on the side of leaving in too many words than removing too many words.

For (2), in the same way that compressors, bandpass filters, gates, etc., allow users to specify attack and release time values, it'd be nice if we could specify attack and release time values for the filler word remover. That might help to smooth out some of the choppiness when removing filler words. (I suppose this might imply that the remover would replace filler words with silence rather than chopping them out. Maybe that's a preferable way of doing things, though, since it gives people the opportunity to remove those silences later with the tool that chops out word gaps, and/or it gives them space to manually blend the waveforms that come before and after the removed filler word.)

Thanks!

YES. Same with repeated words. The first screenshot is the automatic cut and it sounds like the speak says "regions is" which is incorrect and only because of the weird auto cut. The second screenshot shows her saying "is" 3 times, and if the first is was kept because the waveform is joined with the words, it sounds great.

Tl;dr, remove filler words and repeated words only when the waveform goes to zero for a clean cut. With all others that can't make the clean cut automatically, put a marker for review.

and if you HAVE to do it this way, put a marker for manual review of each cut done automatically for accepting or rejecting. It's not easy to hunt for and listen to each each auto edit, expand it just in case there's a better edit to be made, make that better edit, listen again, and finally move on. It takes MORE time.

(Capitals for emphasis, not anger) cc @Dan Reyes-Cairo

yes, this would be great! I never use the feature because it creates harsh chops.