What if a voice assistant waited through your pause rather than cutting in? On October 23, 2026, Block unveiled Dessa, an experimental AI model that listens and talks simultaneously.
The team says Dessa learned from recordings of human conversations, without training on text or starting with an existing language model. Its demos suggest more flexible conversation. But the company has not released independent performance results or made the model widely available.
Key Highlights
- Dessa is a full-duplex, speech-to-speech model announced on October 23.
- Block says it trained the system on audio alone using one NVIDIA GB200 NVL72.
- The model reportedly runs locally on a MacBook Pro M4 Max or better.
- Safety researchers are being offered limited access; public download and API details remain unavailable.
What Did Block Reveal About Its New Voice AI Model?
A small research team within Block, led by Ragavan Thurairatnam, developed Dessa. The group traces its roots to a Toronto AI company of the same name, acquired by Square in 2020.
On the official Dessa project page, Block released recordings of unscripted conversations. In one, the model laughs midway through a sentence. In another, it raises its voice after being asked to speak louder. It can also offer brief acknowledgements while someone continues talking.
Jack Dorsey highlighted the launch in an October 23 X post, while Block’s official account shared the recordings. These are not controlled reliability tests. Some exchanges have also been criticised for awkward responses and mistaken assumptions about the speaker.
How Does Dessa Listen and Speak at the Same Time?
Conventional voice assistants often wait, transcribe speech, process text, and produce an audio reply. That chain can miss a person’s tone or treat a hesitation as the end of a question.
Dessa takes incoming audio and produces spoken replies directly. Block describes the system as full-duplex, meaning its listening channel stays active while it talks. That allows behaviours normally seen in face-to-face conversation:
- Interruptions: A caller can correct the assistant without waiting for it to finish a long answer.
- Backchannels: Short responses such as “yeah” can acknowledge speech without taking over the conversation.
- Vocal cues: Pauses, laughter, and changes in volume are potentially available to the model rather than being flattened into a transcript.
This approach is not entirely new. OpenAI’s GPT-Live-1 also supports full-duplex dialogue. Dessa’s distinctive claim is its speech-only training, although outside researchers have yet to verify how that affects accuracy or emotional awareness.
What Could Change for Digital Assistants and Customer Support?
Consider a customer phoning about a failed payment. A turn-based bot might keep reading an explanation after the customer says the problem has already been solved. A full-duplex assistant could hear the interruption and change course sooner.
Other possible uses:
- Customer helplines: Conversations could become less frustrating when people change their question mid-answer.
- Language practice: Learners could pause, restart a sentence or receive brief encouragement without triggering an unwanted lecture.
- Accessibility: Voice-first interfaces may become easier to navigate for people who rely on spoken commands rather than screens.
These are possible applications, not deployed Dessa products. Handling emotion is especially sensitive: a system that sounds sympathetic may still misunderstand distress, sarcasm or consent. Research on existing voice systems has documented gaps between recognising a vocal cue and acting appropriately on it.
Can Anyone Download Dessa, and What Is Still Unknown?
Not yet. As of October 24, Block presents Dessa as a research preview, with early access intended for a small group of AI safety researchers. Its launch page mentions a wider waitlist, while download links are not active.
Block says training used one NVIDIA GB200 NVL72 rack, and claims inference runs faster than real time on an M4 Max MacBook Pro or stronger hardware. Independent benchmarks remain unavailable.
The company has not published the parameter count, detailed training dataset, supported languages, model weights, API pricing or a general release date. Running locally could eventually reduce dependence on cloud processing, but that does not establish what data an eventual application would store or transmit.
For now, the recorded conversations are worth examining without treating them as proof of everyday performance. Follow Block’s official Dessa updates for access details and technical evaluations.
FAQs
What Is Dessa Voice AI?
Dessa is Block’s experimental speech-to-speech model. It processes conversational audio directly and is designed to listen while talking, rather than waiting for strict speaking turns.
Who Developed the Dessa AI Model?
A small research team within Block built Dessa under Ragavan Thurairatnam. The project draws on staff from the earlier Toronto startup acquired by Square in 2020.
Can Dessa Run on a Personal Laptop?
Block says it can run locally on a MacBook Pro M4 Max or better. Public access and independently verified performance figures are not available yet.
Is Dessa Better Than Existing Voice Assistants?
There is no verified ranking. Block has released demonstration recordings, but independent benchmarks and public testing are still needed to judge accuracy, interruptions, and long conversations.
When Will Dessa Become Available to Everyone?
Block has not announced a public release date. It is inviting select safety researchers first, while broader access, downloadable model weights and commercial terms remain unspecified.
Discover the AI Breakthroughs Changing Our World!
1. Can Government AI Systems Stay Secure?
Explore seven cybersecurity risks India must prepare for as AI enters government systems.
2. What Can Claude Voice Mode Control?
Discover Claude’s voice features, Opus and Sonnet capabilities, and app control options.
3. Could AI Preserve India’s Tribal Languages?
Uncover how Adi Vaani uses AI translation to support India’s tribal languages.
4. Why Are Tech Workers Luck Maxxing?
Find out why tech professionals are embracing luck maxxing and reconsidering their careers.
5. Can Your Face Reveal Biological Age?
See how AI analyses facial features to estimate biological age and potential health risks.
