Skip to content
Neural Daily

ModelsNews

Block's Dessa voice model learned from speech alone

🇭🇺 Olvasd el magyarul →

Block's research lab unveiled Dessa, a full-duplex speech-to-speech model trained without any text, only on recordings of human conversation. It is a research preview with a waitlist.

By: Mikola Máté Patrik · · 2 min read

Dessa logo by Block: a light green looping line forming a circle on a black background
Image: Block · Press kit · Official Dessa page social image (og:image) · Source

What is it?

On October 23, Jack Dorsey's Block unveiled Dessa (Block), a speech-to-speech model: audio in, audio out, without converting speech to text along the way. Block says it was trained from scratch on human conversation audio only, with no text and no pretrained models.

What it can do

  • Full-duplex: it listens and talks at the same time, knows when to cut in and backchannels (“yeah”, “got it”).
  • It hears and expresses emotion, including laughter, sarcasm and sighs, which get lost when speech is flattened into text.
  • Inference runs faster than real time, so there is no wait for a reply.
  • It runs locally on a MacBook Pro M4 Max or better.
  • Block says it learned language from hundreds of times fewer words than LLMs, and it was trained on a single NVIDIA GB200 NVL72 system.

The demo clips are live conversations with subtitles added afterwards; one shows Dessa talking to OpenAI's GPT-Live-1. The Verge notes the chats can be awkward at times: in one clip the model first mistakes the speaker for someone named “Ronaldo”. In the exchange with GPT-Live-1, though, Dessa comes across as the more natural of the two.

Who built it

A small research lab inside Block, founded by former members of the Toronto deep learning startup Dessa. The team reports directly to Jack Dorsey and is led by Ragavan Thurairatnam.

Can you try it?

Not yet. There is a waitlist, but no API, weights, pricing or benchmarks. Block is giving a small group of AI safety researchers early access to look for risks the team hasn't anticipated.

Why it matters

Most voice AI today chains speech recognition, an LLM and speech synthesis, which adds latency and loses tone and emotion. Dessa shows a conversational model can be trained without text, on relatively modest compute. How it holds up on factual knowledge and reliability remains an open question until independent tests arrive.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles