I second Grzegorz's feedback. I wish it were possible to know after the fact which model has been used and/or be able to select the model before initiating the transcription.

In my case, the automatic transcription is so bad that it took me 60 minutes to do a first pass correcting the transcript of a 50-minute podcast episode. And I still have to do a second pass for the bits of audio that haven't been transcribed for some unknown reason.

In comparison, I uploaded the same raw recording into Riverside and I noticed that the initial transcription took a few minutes longer than in Descript. However, 100% of the audio got transcribed and the quality of the transcript was so much better; it only took me 10 minutes to correct the handful of transcription errors from their model.

Hi Grzegorz!

I'm sorry you're having issues with transcription. The team has been making a number of changes to improve the transcription experience in the last few months, with plenty more to come.

One of the new features we recently introduced is the option to change your Transcription Provider in the General Settings tab of your Descript account. By default, you should be on Automatic, which would use "Scholar", which we've determined is the best provider in most cases for most people. You can also try "Daredevil" (which is Whisper under the hood). We've seen Whisper is far more prone to hallucinations in broad testing, and a slightly lower overall word error rate, despite being faster and slightly better at proper nouns.

We're working to add new models over time, so if you want what Descript thinks is generally the best you can set to automatic, and we'll keep you on our recommended provider. But you may find that one of the others works better for your specific content and want to stick with them.

Aside from the providers themselves, we continue to invest our own logic to improve the transcriptions and transcript editing experience.

Thank you for the feedback!