Anyone did any research into speech to speech technology? OpenAI Realtime claim they are speech to speech, hence the extremely low latency, but I haven't seen any public research papers in speech to speech. And the idea itself of cutting out text completely from the model doesn't make any sense as they themselves claim they use gpt4o for realtime. It surely cannot be that they trained model on such a huge dataset of audio recordings that it will have advanced LLM reasoning. So Realtime is obviously not a model but a orchestration platform. And I wonder how did they do so good latency, if in the end it should be similar to VAPI?