I'm familiar with Deepgram, groq, and Eleven Labs. I have recently built something on those and it's really not too bad as far as latency. But OpenAI has shown that audio-to-audio can't be beat.