Dessa: Block's conversation model trained only on speech
Dessa: Block's conversation model trained only on speech
Block's AI research lab trained a speech-to-speech model from scratch on recorded human conversation, with no text at all. It is full-duplex: it listens and talks at the same time, with overlaps, laughter and "yeah" backchannels. The samples sound very human.
Rufus's note (2026-10-25): Very realistic sounding (listen to the samples on the page) and trained only on human speech. Useful in itself, but also a demo of what you can now do: trained on one rack, not a cluster. — Rufus Pollock
Links
- Page with audio samples: https://block.xyz/dessa (released 23 Oct 2026)
- More recordings: Ragavan Thurairatnam on X
- Discussion: X trending summary
Overview
Most voice assistants convert speech to text, run a language model, then convert back. Dessa skips text entirely and learns turn-taking and tone straight from audio. That is why it can interrupt and be interrupted naturally.
It was trained on a single NVIDIA GB200 NVL72 rack rather than a large cluster, and runs locally on a MacBook Pro M4 Max or better, faster than real time. It is as much a demo of what a small lab can now do as a product.
Access
No weights, licence or paper yet ("launch links are coming soon"). There is a waitlist, and Block is giving a small group of AI safety researchers early access. The obvious worry raised on release: voice that convincing makes phone scams easier.
See also
- moc-tts-voice-models: text-to-speech models (Dessa is speech-to-speech, so it is not listed there)