Block published a Dessa speech-only full-duplex research preview with audio demonstrations.
Block previews Dessa, a speech-only full-duplex conversation model
2026-10-23Frontier watchWATCH
Block previewed Dessa, a speech-to-speech model trained from scratch on human conversation without text or a pretrained model. It listens and speaks simultaneously and runs faster than real time, according to the team. The page provides selected audio demonstrations but no weights, code, API, benchmark protocol or independent evaluation.
Availability: selected audio demos only; no weights, code, API or benchmark protocol
What this means for you
For voice-agent builders, speech-only training could preserve timing, overlap and emotion that text intermediates discard. Without a testable artifact or systematic evaluation, the demo cannot yet support an architecture decision.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Training | trained from scratch on human conversation audio; publisher says no text or pretrained model |
|---|---|
| Interaction | full-duplex speech-to-speech with simultaneous listening and speaking |
| Infrastructure | publisher says training used one NVIDIA GB200 NVL72 |
| Availability | selected audio demos only; no weights, code, API or benchmark protocol |
| Time Basis | the official page establishes October 23 at date precision; Hacker News submission at 2026-10-23T19:09:54Z is an external upper bound only |
Technical impact 4
Engineering value 3
Adoption signal 1
Maturity 1
Verification cost 5
Risk 3