Dessa

Block's full-duplex speech-to-speech model trained from scratch on human conversation, small enough to run on a high-end laptop.

Block · Generative media (image, video, audio, music, voice/avatar generation) · model

Published in Feed

Overview

Block (Jack Dorsey's company) introduced Dessa on October 23, 2026, a speech-to-speech model trained from scratch on conversational speech rather than text or a pretrained language model. It is full duplex, so it can listen and talk simultaneously, interrupt and acknowledge without waiting for turns. Block says it runs locally on consumer hardware such as a MacBook M4 Max or better. Access is limited to a small group of AI safety researchers plus a waitlist; Block has not published parameter count, weights or standard benchmarks, and its demo compares Dessa against OpenAI's GPT-Live-1 without that team's involvement.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Full-duplex speech-to-speech: listens and speaks at once
  • Trained from scratch on human conversation, not text
  • Runs locally on a MacBook M4 Max or better (company claim)
  • Early access limited to AI safety researchers; public waitlist
  • No parameter count, weights or benchmarks published

Use cases

  • Natural voice assistants that handle interruptions and backchannels
  • On-device voice interfaces without cloud audio upload

Pricing

unknown; not publicly available, waitlist only.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
23 Oct 2026
Record updated
28 Oct 2026
Record last checked
28 Oct 2026

Explore the Feed

Other entries filed under generative media (image, video, audio, music, voice/avatar generation); inclusion does not imply equivalent features.

  • Amazon Nova 2.5 Sonic

    Amazon's generally available real-time speech-to-speech model in Bedrock, with a 256K context window and asynchronous tool calling.

  • Grok Voice 3.0

    xAI's speech-to-speech model in the Voice Agent API, priced at $0.08/min and scoring 86.9 on Artificial Analysis' Speech-to-Speech Index.

  • Helix World

    Real-time interactive world model from Noiz AI and HKUST that generates explorable video with spatial stereo audio from one image.

Browse the Feed