According to @_avichawla, Deepgram Flux TTS keeps acoustic state, enables sub‑200 ms first audio, and logs interruptions for coherent voice agents.