Whisper's turbo model beat large on noisy audio, in practice

Building the Video Translator , the assumption going in was that OpenAI Whisper’s larger model would transcribe more accurately. In practice, on clips with loud background music, the turbo model did noticeably better than large. It seems to tolerate messy audio better, rather than just being a faster, lower-quality option. Separating vocals from background music first (with demucs) didn’t help as much as expected either; Whisper handled the mixed track better on its own. Worth trying turbo first rather than assuming bigger is more accurate.