Block (the Cash App people) just showed a voice model that learned to talk only from people talking. No text, no pretrained model, just recorded human conversation.
It's called Dessa. Full duplex, so it listens and talks at the same time, and it can laugh and sigh too. Runs on a MacBook Pro with M4 Max, faster than real time. For now only a research preview with a waitlist.
Their lab lead called it "a GPT3 moment for speech". Big words :) but for me the exciting part is local. No server, no round trip, no awkward pause before the voice answers.
Article here if you're curious: https://www.aichatdaily.com/ai-models/block-unveils-dessa-full-duplex-voice-model-trained-only
Get an email when Kristjan publishes