Have transcript learn from manual edits
M
Matty Dalrymple
The Script Editing Basics tutorial suggests that this should happen automatically (i.e., when I edit "cotton tent" to "continent" in one place in the tutorial transcript, it makes that correction throughout), but my actual experience doesn't reflect this (I have to change "drafted digital" to "Draft2Digital" repeatedly.)
Jeremy Andrew Davis
YES! Why does this not happen? I've been using Descript for years and it has yet to learn basic disability terminology that I use all the time in my videos. You would think it would at least learn my user's preferences if not take that data and inform it's overall AI. I spend so much time changing the same word over and over in a single script. This needs to be implemented months ago.
K
Karen Schouest
I wonder if the reason it can't learn from repetitive editing/correcting is because Descript doesn't have control over the internal workings of the AI engine, since it belongs to Rev.ai. Maybe that's something that would have to come from the Rev.ai developers, and it's not on their priority list. And perhaps the AI engine for Otter.ai, along with all the other components, actually belongs to Otter and can thus be enhanced at their whim. Just a theory. It seems like it would be something that Rev.ai would want to do to stay competitive with its product.
Interestingly, I have run some audio through the actual Rev.ai website, and then run the same audio through Descript, and the transcription output was different, with the Descript transcript being better quality. I've also put the same audio through Otter.ai and Descript, with similar results of the Descript transcript being a better quality for my purposes. Not by a landslide, mind you, but enough to be noticeable and keep me coming back to Descript as my AI engine of choice. So Descript is obviously doing
something
that improves the output. If only that "something" could include paying attention to our corrections and learning as it goes.Charles-Alexandre Roy
I just want to second dan's comments after having used Otter.ai for a while as well. The three features that are most lacking relative to Ottar.ai are:
- Continuosly updating the model based on manual edits (and then reapplying those updates everywhere).
- Updating the speaker model after manually changing them (and then applying those updates everywhere)
- Learning and using account-wide custom vocabulary.
C
Christopher Mackay
I’m transcribing art instruction videos, and I run into this constantly. “BiC pens” comes out as “big pens”, “Moleskine” as “mole skin”, etc. I don’t expect Descript to transcribe these super-nuanced homonyms accurately (most people wouldn’t, either), but once I’ve corrected it, it feels like I shouldn’t have to do it over and over.
I know. Easy to describe; next to impossible to fix. :)
D
David Friedman
Christopher Mackay: I'm just adding a note here for people who might not know about the Transcription Glossary. For me, a lot of this has been solved by adding commonly-misunderstood words to the glossary. I'm curious if you've tried that for your most common art instruction terms.
C
Christopher Mackay
David Friedman: Nope! Hadn’t noticed this command — will check it out next time I’m editing a transcript. Like others, I’d thought this was something Descript “learned” automatically (but was failing to do for whatever reasons).
C
Christopher Mackay
David Friedman: I’d also like to be able to add “stop words” to the glossary (anti-glossary?). I’ve seen some unprintable/outrageously offensive terms come up in transcripts that WERE NOT in the source audio. Relying on catching these manually is a sobering thought.
S
Shannon Wedge
David Friedman: I find it totally ignores at least 1/4th of the things I have in my Transcription Glossary. Despite hearing the same things repeatedly by mostly the same voices, it still consistently gives National Online as "nationally," calls Joni "Johnny," and stubbornly outputs biopsychosocial as bio-pyscho-social.
S
Sat P
David Friedman: I never knew the Transcription Glossary existed! Thanks for pointing this out, will take a look today.
d
dan
Oof, I assumed this was already built in.
I've been using otter.ai for transcribing audio. Their transcripts continuosly update as you make edits. There's also a Rematch Speakers function that is really good and reruns as the speakers are definined. The speakers carry over transcript to transcript, it's pretty accurate after a few rounds.
They also support account-wide custom vocab and names of people. 30 second video of that: https://www.youtube.com/watch?v=8HuA9uZ8iRQ
I guess the next step is seeing if I can use their transcripts in descript a la https://help.descript.com/hc/en-us/articles/360042210632-Import-Transcript
Hey Andrew Mason what do you say we kick off 2022 right and bump this mamajama up to Under Review? :D
Murray Robinson
dan: Andrew Masonspeech-to-text Otter .AI is far better at transcribing speech than Descript is. I would love it if Descript used it as the speech-to-textspeech to text tool rather than whatever they are using now
M
Michael Ferguson
Yes!
Gabriel Serafini
I would like also for (with our opt-in permission) our manual corrections to filter back into the transcript learning side, so that the quality of fresh transcriptions improves over time. Ideally if we manually correct a transcript (for even common words / phrases) that the auto-transcription got wrong, eventually I would like our (many) inputs to improve the underlying engine so that if we uploaded an identical clip later, it would get it closer to 100% correct.
S
Sat P
Gabriel Serafini: Agreed. I would love to understand how the transcription should get better over time. It feels as though it doesn't get better at all.
Ron Stauffer
Yes! Please add this... every single time I say my company name on video, Descript misspells it in the transcription in every instance and it would be nice if it could learn from this and fix them all.
Frameworks
Agreed. Furthermore, as a first step since Descript already offers a "Remove filler words...", perhaps engineers there could quickly adapt that feature to offer an additional feature to "
Apply common corrections...
". This new feature would track, record and apply (on your choosing) a list of your most frequently performed word replacements or phrase corrections. That would go along way to helping Descript users increase productivity during transcription activities.
B
Blaine Moore
I have a list of terms I have to go in and bulk edit every week.
S
Shannon Wedge
Blaine Moore: Me too
Load More
→