Why not Whisper on device, or Apple's Speech framework? Why send audio to OpenAI at all?