Dessa (Block)

What Dessa (Block) does

Research-preview voice AI from a small AI lab inside Block (Square, Cash App), reporting to Jack Dorsey and led by Ragavan Thurairatnam, co-founder of the Toronto startup Dessa that Block acquired. Dessa is a full-duplex speech-to-speech model: audio in, audio out, listening and talking at the same time so it can interrupt, laugh mid-sentence and say 'yeah' while you talk. Block says it was trained from scratch only on recordings of human conversation (no text, no pretrained models), on a single NVIDIA GB200 NVL72, and that it runs faster than real time locally on a MacBook Pro M4 Max or better. Right now it's a demo page of recorded conversations plus a waitlist; a small group of AI safety researchers get early access.

Good for

• Listen to the recorded demos on block.xyz/dessa (e.g. 'Refund', 'Life advice', the chat with ChatGPT's voice model) and note how interruptions and timing compare with the voice AI you use now
• Join the waitlist if natural-sounding voice matters to you (phone support, voice agents, podcasts) so you hear when access opens
• Use the demos as a benchmark: ask your current voice-agent vendor how they handle barge-in, backchannels and emotion, and compare
• If you work with a university or safety lab, apply for researcher early access on the same page

Honest take

Some claims not independently checked

• But there's nothing to use yet: it's a waitlist only, with no paper, weights, benchmarks, pricing or release date.
• The demos are hand-picked recordings.
• Full-duplex voice itself isn't new (e.g.
• Kyutai's Moshi, OpenAI's and Google's live voice modes); the novelty claimed is training without any text.
• A model that learned facts only from audio may also know less and be harder to steer than text-based assistants.
• UK: the page and waitlist are open worldwide; no UK restriction found.
• The site's privacy notice is from Block, Inc.
• (US).
• Post: about 44K views and 1.1K likes at read.

What we couldn’t verify

• Fact vs hype: Real announcement from Block (block.xyz/dessa, 23 Oct 2026), covered by The Verge (which called some demos 'a little awkward and uncanny'; in one, the model mistakes the speaker for 'Ronaldo') and Unite.AI.
• Every claim is Block's own and unverified: 'trained only on speech, no text, no pretrained models', 'most authentic full-duplex speech model ever built', 'world's first' realistic speech-only conversational AI, trained on one GB200 NVL72, runs on an M4 Max MacBook, and the lead's 'GPT-3 moment for speech'.

Our Hype or Fact? checks

• Partly factBlock's Dessa: a voice AI that listens while it talks 23 Oct

Pricing

Contact sales / book a demo (no public pricing). Not a product yet: no app, API or download, just a public waitlist on block.xyz/dessa and early access by application for AI safety researchers (institutional email required).

Prices as listed by the maker when we last checked (2026-10-23); usually in US dollars. Check the website for current prices.

Latest update

• 2026-10-23: Unveiled Dessa, a full-duplex speech-to-speech model trained only on human conversation audio; demo recordings, waitlist and early access for AI safety researchers. No pricing, API or release date.

Is Dessa (Block) worth it? Quick answers