Dessa Voice: What Block's Speech-Only AI Model Can Do
Dessa voice is Block's full-duplex speech model, trained only on human conversation. See what Dessa voice can do, how it works and how to get access.
By Dessa AI Team · Updated
What Is Dessa Voice?
Dessa voice is a speech-to-speech conversation model from a small AI research lab at Block. Audio goes in and audio comes out: it listens and talks at the same time, and Block trained it from scratch on recorded human conversation, with no text and no pretrained models. Block calls Dessa voice "the most authentic full-duplex speech model ever built." It is a research preview with a waitlist; there is no public API yet.
The lab was founded by former members of Dessa, a Toronto deep learning startup, and reports to Block cofounder Jack Dorsey. Ragavan Thurairatnam leads the team.
Dessa AI, the voice studio you are reading this on, is independent and is not run by Block. The models behind its tools are listed in the Dessa AI FAQ.
Dessa Voice at a Glance
| Dessa voice | What Block says |
|---|---|
| Type | Full-duplex speech-to-speech model (audio in, audio out) |
| Maker | AI research lab at Block, led by Ragavan Thurairatnam |
| Training data | Human conversation audio only; no text, no pretrained models |
| Training hardware | One NVIDIA GB200 NVL72 |
| Runs on | MacBook Pro M4 Max or better, faster than real time |
| Access | Waitlist; early access for a small group of AI safety researchers |
| API, weights, pricing | Not announced |
| Languages | Not listed |
How Dessa Voice Works
Most voice assistants are a chain: speech to text, a text model, then text to speech. Block says text-based pipelines lose information such as emotion, sarcasm, laughter or a frustrated sigh. Dessa skips the text step, so it hears how something is said and not only what is said.
Because the model is full-duplex, it can backchannel ("yeah", "got it") while you talk and, in Block's words, knows "when to cut in." Block also says Dessa voice learned language from "hundreds of times fewer words" than large language models, but it has not published figures for that claim.
What Dessa Voice Can Do
Block's demos are live recordings with no steering; subtitles were added afterwards. In them, the model:
- laughs mid-sentence and gets louder when asked to speak up;
- gasps "oh my god" at good news;
- finishes a person's sentence when they forget a movie title;
- reacts to emotion, sarcasm, laughter and sighs, and changes its intonation and speed;
- holds a conversation with GPT-Live-1, another AI voice model.
These are Block's own demos. No independent benchmark has tested Dessa voice yet, so you cannot compare it with other models on leaderboards. Our best AI voice model comparison lists Dessa as the model to watch for that reason.
Dessa Voice vs GPT-Live-1
GPT-Live-1 is the closest model you can use today: it is also full-duplex. The table shows what each maker has published.
| Dessa voice (Block) | GPT-Live-1 (OpenAI) | |
|---|---|---|
| Full-duplex | Yes | Yes |
| Training | Human speech only, no text | Not described on the model page |
| Public API | No | Yes, Live endpoint |
| Price | Not announced | $0.05 per minute |
| Runs on a laptop | Yes, MacBook Pro M4 Max or better | No, cloud API |
| Independent benchmark | None yet | 83.2 on the Artificial Analysis speech-to-speech index |
How to Try Dessa Voice
Dessa voice is not open to the public yet. To ask for access:
- Open Block's Dessa project page and select Join the waitlist.
- Enter an institutional email address from a university, company or lab.
- Optionally add a Google Scholar profile or blog link.
Block says early access goes to a small group of AI safety researchers, because a model that learned from sound alone is "new territory" and Block wants outside researchers to find risks it has not anticipated. Block says launch links are "coming soon."
To try a full-duplex voice chat in the browser, open Live Talk in the Dessa AI studio; the model behind it is listed in the Dessa AI FAQ. Read how to use Dessa AI for the steps.
FAQ
What is Dessa voice?
Dessa voice is a full-duplex speech-to-speech model from an AI research lab at Block. Block trained Dessa voice only on recorded human conversation, so it hears tone, laughter and sighs directly instead of working through text.
Who made Dessa voice?
A small AI research lab at Block made it. Former members of Dessa, a Toronto deep learning startup, founded the lab, and it reports to Jack Dorsey. Dessa AI, the voice studio, is a separate company and is not run by Block.
Can I use Dessa voice now?
Not yet. Dessa voice has a waitlist, and Block is giving early access to a small group of AI safety researchers. There is no public API, pricing or open weights.
Is Dessa voice free?
Block has not announced pricing for Dessa voice. Early access is by waitlist only.
What languages does Dessa voice speak?
Block does not list supported languages. Its page treats language learning as an open research question.
Which model does Dessa AI use?
Dessa AI is an independent voice studio and is not run by Block. The models behind Live Talk and Text to Speech are listed in the Dessa AI FAQ.
Sources
Every fact about Dessa voice in this article comes from these pages:
- Block: Dessa project page
- OpenAI: GPT-Live-1, pricing
- Artificial Analysis: speech-to-speech index