Block Launches Dessa, a Speech-Only Full-Duplex Voice Model

Block, the company backed by Jack Dorsey, released the conversational model Dessa on October 24, officially describing it as the first speech-only trained AI capable of real-time, human-like conversation. The model was trained entirely from scratch using only human speech data, with no reliance on text or any pretrained models, and supports full-duplex speech-to-speech communication — a release widely seen as a milestone for voice interaction.

Confirmed

• Dessa was trained from scratch purely on human conversational audio, with no text and no pretrained models, reportedly making it the first model of its kind to learn from speech alone.
• It supports real-time full-duplex conversation: it can laugh mid-sentence at something funny, speak up when asked, and carry dialogue with emotion, subtle intonation, pauses, and natural rhythm.
• Dessa runs locally on consumer-grade GPUs, with no need for high-end compute.
• The team has also invited AI safety researchers to study the model.

Why it matters

• The team believes voice is the most natural way to interact with intelligence, so they built the entire model around speech rather than following the traditional "ASR→LLM→TTS" voice assistant pipeline.
• Sharer @ZeroStateReflex called it "the GPT-3 moment for speech," suggesting speech-native models could become a new technological paradigm; @Scobleizer and others described it as "the most powerful voice model." Coming from a team led by Jack (Dorsey) with former OpenAI/Tesla backgrounds, the release has drawn significant attention.

2026-10-24 ~ 2026-10-24 · 7 related posts

• [source] Block Unveils Dessa, a Full-Duplex Speech-Only Model Trained From Scratch With No Text — stochasticchasm · 2026-10-24
• Dessa debuts as a full-duplex speech model trained from scratch on audio alone — dragon_khoi · 2026-10-24
• Jack Dorsey's Blocks debuts Dessa, an AI model trained from scratch on speech only — smtabatabaie · 2026-10-24
• Dessa Debuts as a Full-Duplex Speech Model That 'Feels Like Talking to a Person' — Scobleizer · 2026-10-24
• [source] Dessa: first full-duplex speech-to-speech model trained entirely on speech, runs on consumer GPUs — ZeroStateReflex · 2026-10-24
• [source] Block launches Dessa, an AI voice model small enough to run on a MacBook — Polymarket · 2026-10-24
• Block unveils Dessa: hyper-realistic speech-to-speech model trained on a single GB200 NVL72 — altryne · 2026-10-24