We used @GroqInc to build feycher.com and do real time text-to-speech -> llm -> speech to lip sync. And the best part, it was fast enough that we didn't have to even use streaming to shuttle requests.