Speech-to-text engine: microphones, recording, streaming providers (Deepgram, NVIDIA Riva) with key rotation and failover, batch Deepgram. No UI dependency.