Dessa is a voice AI from a small research lab at Block, Jack Dorsey's company behind Square and Cash App, announced in October 2026. It is a full-duplex speech-to-speech model: it listens and talks at the same time, like a person on the phone. Its big claim: Dessa was trained from scratch on human conversation only. No text, no transcripts, no pretrained language model. It learned to talk by listening.
A live, unsteered conversation with Dessa about its own release. It laughs mid-sentence when something is funny, and speaks up when asked. Recording by Block, from the Dessa page. Contains explicit language.
Why it's different
Most voice assistants are a pipeline of 3 models:
- Speech-to-text turns your voice into words.
- A text model (an LLM) writes the answer.
- Text-to-speech reads the answer out.
Every step throws something away. The transcript keeps the words but loses the tone, the sigh, the laugh, the sarcasm, the hesitation. And the turn-taking is rigid: the assistant waits until you stop, then talks.
Dessa skips text entirely: audio in, audio out.
| Pipeline voice AI | Dessa |
|---|---|
| Hears the words of a transcript | Hears the voice: emotion, sarcasm, laughter, a sigh |
| Waits for you to finish, then answers | Listens and talks at the same time (full duplex) |
| Laughs "on cue" after you stop | Laughs the moment something is funny, mid-sentence |
| Silent while you talk | Backchannels: "yeah", "got it", "ooh" |
| Learned language from huge amounts of text | Learned from speech only, with 100s of times fewer words |
Block's example: say "I didn't say she stole my CUDA code" 6 times, stressing a different word each time. The words are the same, the meaning changes every time. A transcript can't tell them apart.
What the demos show
Block published about 20 live recordings of people talking to Dessa, with subtitles added afterwards. Nobody steers it: Dessa picks what to say, when and how, down to its own voice and personality.
- Timing: it interrupts, takes its turn and finishes your sentence when you blank on a movie title.
- Emotion: at an anniversary surprise it gasps "oh my god" under its breath, then answers with what sounds like teary gratitude.
- Volume and pace: ask it to speak up and it gets louder. Block says nobody taught it that.
- Knowledge: it knows a surprising amount, given that every fact came from raw sound.
- Against another AI: in 2 clips it talks with OpenAI's GPT-Live-1, with fast backchanneling and patience when interrupted.
An anniversary surprise: Dessa's gasp and quiet "OMG". Recording by Block, from the Dessa page.
How it was built
- Data: human conversation only. Block doesn't say how many hours.
- Why that's hard: speech never repeats. The same sentence said twice is different data, and 1 spoken word takes 1,000s of times more bytes than the written word. The model also has to learn accents, pauses, breathing and when to cut in, not only language.
- Compute: trained on 1 NVIDIA GB200 NVL72, a single rack, which is small for a frontier model.
- Speed: inference runs faster than real time, so there is no awkward wait.
- Local: it runs on a MacBook Pro with an M4 Max chip or better.
- Not disclosed: the parameter count, the hours of training audio, and whether the weights will be released.
Who's behind it
The lab is a (very) small team at Block, founded by former members of Dessa, a Toronto deep-learning startup, and led by Ragavan Thurairatnam. It reports directly to Jack Dorsey.
- 2016: Dessa is founded in Toronto, first as DeepLearni.ng, by Eric Kin-Ho Lee, Ragavan Thurairatnam, Stephen Piron and Vincent Wong.
- 2019: Dessa becomes known for RealTalk, a fake of podcaster Joe Rogan's voice convincing enough to make headlines, and for open-source deepfake detection.
- February 2020: Square, now Block, acquires Dessa.
- October 2026: a lab of former Dessa members at Block releases the Dessa model.
The team says it has built deep-learning models since 2012 and generative speech models for over 8 years. For the wider topic of synthetic voices, see Voice Cloning Recording Script.
Why Block built it
The lab gives 3 reasons:
- Curiosity: babies learn to talk without seeing a written word. The team wanted to know if a model could too.
- Better interaction: speech is the most natural interface to any intelligence, and pipeline voice AI wastes it.
- Safety: AI is gaining power and independence while staying nearly deaf to emotion. Block's example: a person in serious pain, and a voice AI that has no idea and keeps going.
It also opens research questions: how efficiently a speech-only model learns language and facts, how fast it learns a new language, and what in-context learning it can do.
Availability
- Waitlist on the Dessa page. No release date, no price and no API yet.
- Safety researchers can apply for early access, to find risks the team hasn't anticipated.
- Weights: not announced as open.
Compared with other full-duplex voice models
| Model | From | Trained on | Open |
|---|---|---|---|
| Dessa | Block | Human speech only, no text | Not announced |
| Moshi | Kyutai | Speech and text, 7B parameters | Yes, CC-BY 4.0 |
| CSM | Sesame | Conversational audio, Llama-based | 1B model open, larger models closed |
| PersonaPlex | NVIDIA | Built on Moshi, real and synthetic dialogue | Research release |
The others start from text or from a model that already knows text. Dessa's novelty is that it doesn't.
Limits and open questions
- Demos, not a product. The clips are curated recordings. Nobody outside Block has tested it yet.
- Facts from sound only. It knows a lot, but how reliably it gets facts right, compared with a text-trained model, is unknown.
- No tools, no text. Nothing in the release says it can read documents, call APIs or write. It is a conversation model.
- Realism cuts both ways. A voice that laughs, gasps and sounds moved is easier to trust, and to misuse. The same team made the RealTalk voice fake in 2019, which is part of why it asks safety researchers in first.
Further Reading
- Kyutai, makers of Moshi, the open full-duplex model
- NVIDIA PersonaPlex, a full-duplex model built on Moshi