Speech-to-text on a fanless mini PC

A small multilingual speech model transcribes my meeting recordings at a real-time factor of about 0.35 on a 4-core fanless box using int8 weights.

Roughly 70% of the time is the encoder; batching 30-second chunks and pinning threads to physical cores gained more than any model swap. Thermal throttling after ten minutes is now the real limit.