There's a better parallel/batching that works on the 30s chunks resulting in 40X. From HF at https://github.com/Vaibhavs10/insanely-fast-whisper

This is again not native PyTorch so there's still room to have better RTFX numbers.