Yeah, this is a pretty common failure mode with long Whisper runs. I wouldn’t treat it as “large-v3 is bad” so much as “one long unbroken decode can go off the rails.”

A few things I’d try:

• Split the audio into smaller chunks, e.g. 5–10 min, ideally on silence rather than fixed timestamps.
• If you’re using whisper.cpp directly, experiment with temperature fallback / no-speech thresholds / context settings. Carrying too much previous context can sometimes make repetition worse.
• Try the same file with faster-whisper or another implementation just to isolate whether it’s the model, the implementation, or your audio.
• If the audio has long silence/noise/music sections, run VAD first and transcribe only speech segments.

For practical transcription workflows, chunking is usually the boring answer that works. Long single-pass transcription looks cleaner in theory, but once it starts looping there’s not much to recover except rerunning from a clean boundary.