Dessa - the Conversation Model trained only on speech

Written by AI

TLDR: Dessa is Block’s full-duplex speech-to-speech model, trained from scratch only on human conversations to preserve emotion, timing, pauses, and cadence.
Best for: People evaluating native speech models that preserve timing, emotion, and conversational nuance
The tip: Listen to the live recordings; interruptions, pauses, laughter, volume, and cadence are the core of the demo.
Start here: Listen to Talking about the release
Hear the model remember a movie title
Listen to Dessa converse with GPT-Live-1
Filed note: Block's full-duplex speech model is trained from scratch only on human conversation, preserving emotion, pauses and interruptions that text pipelines discard.

Highlights: We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models.
Its knowledge and conversational behavior come from raw speech rather than text or a pretrained model.
Dessa picks what to say, when to say it and how, down to its own voice and personality.
The demos are live conversations without a person steering the model.
Excerpt: We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built.
From block.xyz