[bensyverson]

There are now realtime “speech-to-speech” models [0]. I believe they skip text to streamline the architecture.

[0]: https://openai.com/index/introducing-gpt-realtime/