Dave on Nostr: I think it's totally possible. On device transcription is already solid enough. Then ...

I think it's totally possible. On device transcription is already solid enough. Then you profile the mp3 waveform and look for anomalies that correspond to what the model thinks is an inserted advertisement from the transcript text. The heuristics of that can get you accuracy high enough that most people wouldn't care if one slipped through every now and then.

Adam proved it works in real time on our last board meeting.