Dessa debuts as a full-duplex speech model trained from scratch on audio alone

dragon_khoi · x · 2026-10-24

Jack announces Dessa, billed as the most authentic full-duplex speech model ever built — trained from scratch purely on human conversation, with no text and no pretrained models.

Key points:

Speech-only training: reportedly the first AI you can hold a realistic live conversation with that learned only by listening.
Truly full-duplex: it laughs mid-sentence when something is funny and gets louder when asked, rather than waiting for you to finish.
The team argues pipeline-based voice AI that converts speech to text discards emotional information, making for worse and potentially unsafe interactions.
Open research questions include how efficiently such models learn language, facts, and in-context behavior.

A waitlist is open, with live recordings of unscripted conversations as demos.

Original post →