OpenAI insists it's not launching a search engine nor GPT-5 on Monday

Rumor has it they're unveiling some sort of conversational, pure-audio model (for ex, leaked code makes reference to some kind of phone call capability). The current ChatGPT app already has great speech recognition thanks to their Whisper transcription model, and it responds with lifelike vocal synthesis, but the content gets transcribed to text in the middle both ways in order to get processed by GPT-4. An audio-only model would be able to process speech directly and respond in kind, applying the conversational interactivity of text language models to the kind of raw-waveform generation that powers music synthesizers like Jukebox and Udio. It might also imply the ability to recognize and respond to the emotional content of speech -- the speech generator from ElevenLabs, for example, is already spooky-good at being able to imbue spoken text with whatever emotional qualities the text itself conveys. If such a model were good enough and fast enough, it would seem eerily like talking to an actual person. This could be revolutionary for agent-like models -- imagine Google Duplex on steroids. And there's a reason a lot of leakers draw comparisons to Her.