Dessa

Dessa: Full-Duplex Speech Model Trained Only on Speech

Dessa is a speech-to-speech conversation model from Block's AI research lab, trained from scratch on recordings of human conversation, with no text and no pretrained models. It listens and talks at the same time and is available as a research preview, with a public waitlist and early access for AI safety researchers.

Key Features

• Audio in, audio out: a true speech-to-speech model, with no transcription step between what the user says and what Dessa replies.

• Full-duplex conversation: listens and speaks at the same time instead of waiting for the user to finish a turn.

• Trained only on speech: learned language and conversation from human conversation audio alone, without text data or a pretrained language model.

• Natural turn-taking: learned from real conversations when to cut in and when to backchannel ("yeah", "got it", "ooh") to show it is following.

• Emotion and tone: hears and produces emotion, sarcasm, laughter, sighs and a voice cracking with sadness, cues that a speech-to-text pipeline drops.

• Own voice and personality: in the published demos, Dessa chooses what to say, when to say it and how, including its voice, with no steering.

• Faster than real time: inference runs faster than real time, so replies come without a pause.

• Runs locally: runs on a MacBook Pro with an M4 Max chip or better.

• Single-rack training: trained on one NVIDIA GB200 NVL72 system.

Use Cases

• Speech model research: studying how a model learns language, facts and new languages from audio, and what in-context learning it can perform.

• AI safety research: early access for safety researchers to probe the risks of a model that learned to talk from sound alone.

• Expressive voice conversation: live conversation that keeps emotion, timing and interruptions, as shown in the recordings on the Dessa site.

• On-device voice AI: local, real-time conversation on high-end Apple Silicon laptops.

Availability

Dessa is a research preview. Block has not announced an API, pricing or model weights. A public waitlist is open, and AI safety researchers can request early access with an institutional email.

Getting Started

• Website: https://block.xyz/dessa

Dessa gives researchers an early look at a conversation model that learned to talk only by listening, with full-duplex timing, native emotion and local inference on a laptop.