Hacker News
Dessa: the most authentic full-duplex speech model ever built
AIFirst seen 477 points180 comments
Dessa is a new voice AI model that learned to talk only by listening to human speech, with no text. It can hold live, back-and-forth spoken conversations that sound more like talking to a person.
What happened
The team behind Dessa released a speech model that they say is the most authentic of its kind. Dessa listens and talks at the same time, like people do. It picks what to say, when to speak, and how to sound. The team shared short recordings of people talking with it, plus a demo where it talks with another AI, GPT-Live-1.
Key facts
- Dessa was trained from scratch on human conversation only, with no text and no pretrained models.
- The team says it learned from hundreds of times fewer words than large language models (LLMs) use.
- It runs faster than real time, so there is no awkward wait.
- It runs locally on a MacBook Pro M4 Max or better.
- The team says it is the first AI you can have a realistic live conversation with that was trained only on human speech.
Why text is a problem
Most voice AI first turns your speech into text. This throws away things text cannot carry, like tone, sighs, laughter, and pauses. The article says this makes voice AI feel slow, stiff, and annoying. Dessa skips the text step and works with sound directly.
Why it matters
Dessa could make talking with AI feel more natural. It can hear feelings like sadness or sarcasm, and it can laugh or say 'yeah!' at the right moment. The team also says it opens new research questions, such as how well speech-only AI learns language and facts compared with text-based AI.
What to watch
- More recordings will be shared on X soon.
- How well Dessa works outside these demos, since the team showed only selected clips.
- Whether other researchers test how speech-only models learn languages and facts compared with text models.