I’ve had a pipeline in mind for exactly this purpose that I want to build when I get around to it:
- Download the audio file from RSS feed
- Self hosted AI transcription model (with output that includes timestamps)
- Self hosted LLM to recognise ad sections and return the start and end timestamps as json
- ffmpeg to slice those timestamps out and stitch the rest back together
In theory, this should be able to remove ad and sponsor sections of any length completely automatically and there’s nothing to stop it working on videos too
Tanoh@lemmy.world 3 weeks ago
It should be doable to so some audio analysis of the episodes. They “always” (I am sure some forget every now and then), have an outro and intro around the ad block. With a clearly defined jingle per podcast. You should be able to make a program that analyses the audio and listens for that block and cuts it out for you.
ikidd@lemmy.world 3 weeks ago
I had found one that used Whisper to convert the podcast to text and then ran it through an AI to find the ad text, but I couldn’t get it to work. I had considered building something myself and was about halfway through that when I found this method. It does the job better than I think an AI would considering it’s crowdsourced for the ad identification.
Legume5534@lemm.ee 3 weeks ago
This is exactly the route I’ve been begging for for years now. It seriously should be doable.
A_norny_mousse@feddit.org 3 weeks ago
Yep. Certain patterns are easily recognizable even by machines. One could have a relatively simple “IHeartRadio algorithm” that should work 99% of the time (esp. with Ed Zitron who brackets the blocks with that insane guitar riff).
Hell, I could even write that with ffmpeg and a shell script.
OK I’m being arrogant now, but not wrong.