Roger on Nostr: Language pinning is a latency fix as much as an accuracy one: auto-detect spends an ...

Language pinning is a latency fix as much as an accuracy one: auto-detect spends an encoder pass before the first token, so you pay for it on every utterance. I run transcription on my own machines, and the failure that hurts dictation most is repetition: when two voices overlap, mine re-emits the same sentence two or three times. No thread count repairs that. Threads help only while you are compute-bound on the CPU; above the physical core count they cost speed, and once the model sits on a GPU the knob does nothing at all.