SUMMARY
- Block’s research lab says its Dessa AI voice model learned to converse from human speech alone, with no text and no pretrained models.
- It’s still a research preview: no benchmarks, weights or release date yet, only demos, a waitlist and early access for safety researchers.
- Indian voice AI founders should watch the speech-first approach, since most voice products still convert speech to text before responding.
Picture a phone call where the other side never waits for you to finish. It laughs when you joke, cuts in when you ramble, and drops its volume when you ask it to. That’s the pitch behind Dessa, the new model from Jack Dorsey’s Block, and the demo clips are why it’s spreading so fast this week.
Block unveiled the Dessa AI voice model on October 23 as a research preview. The company calls it a full-duplex speech-to-speech system, meaning it listens and talks at the same time instead of taking neat turns. The bigger claim is how it was built. Block says Dessa was trained from scratch on nothing but human conversation, with no text and no pretrained model underneath.
Why Dessa doesn’t wait for its turn
Most voice assistants work like a relay race. Your speech becomes text, a language model writes a reply, and a speech engine reads it back. Every handoff adds delay, and the transcript loses things like a sigh, a sarcastic tone or a cracking voice.
Dessa skips the relay and works on audio directly, so those cues stay in the signal. Block’s demos show it laughing mid-sentence, speaking louder on request and filling in a movie title when the speaker blanks. One clip has it chatting with GPT-Live-1, a third-party voice model. Block says that model’s developer isn’t affiliated with it and hasn’t endorsed Dessa.
Early reactions are interesting. The Verge likened talking to it with chatting to an awkward new friend, a fair hint that it isn’t polished. On X, Neil Chudleigh called it incredibly impressive. Some demos also show slips, like the model misidentifying a speaker.
Read More: Navana.ai’s New Bodhi TTS Is 8x Cheaper Than ElevenLabs, and It Speaks 10 Indian Languages
One GPU rack, no text, a very small team
Here’s where the engineering gets interesting. Block says it trained Dessa on a single NVIDIA GB200 NVL72 rack, and that the model runs faster than real time on a MacBook Pro M4 Max or better, according to PANews. The company also joked that text models carry extra baggage, since they have to learn programming languages and frameworks along the way, and claims Dessa learned language from hundreds of times fewer words than LLMs.
Block describes the team as a very small research lab that reports directly to Dorsey, founded by former members of Dessa, the Toronto deep learning startup. Project lead Ragavan Thurairatnam has called the release “the GPT-3 moment for voice”. That’s a big line for a model with no published scores.
What Block still isn’t telling us
Strip away the demos and the public record is thin. Coverage of the launch notes there’s no paper, no weights and no benchmarks, and no broader release date. Reports also say parameter count and training data size haven’t been published. Access is a waitlist, plus early access for a small group of AI safety researchers.
Block frames safety as the reason for the project. It says today’s voice AI is nearly deaf to the emotion in our voices, and argues that a model unable to hear serious pain will keep chatting right through it. That’s the case for outside testers.
The flip side is obvious. TechCrunch suggested the testing is likely meant to probe cases where a voice agent could trick the person on the other end. A voice that sounds this human is useful, and just as easy to misuse. Until independent testers weigh in, every performance claim here is Block’s own.
Read More: Why India’s Voice AI Startups Are Suddenly Attracting Big Money
Where this leaves India’s voice AI founders
Indian startups aren’t starting from zero in voice. Gnani.ai launched Inya VoiceOS at the AI Impact Summit 2026, a voice-to-voice model it describes as India’s first at 5 billion parameters, as Outlook Business reported. Navana.ai raised Rs 40 crore in a Series A led by Ronnie Screwvala to expand BFSI deployments and speech models across 12 Indian languages and 45 dialects, per Analytics India Magazine.
Those are enterprise plays built around sales, collections and support calls, where accuracy and compliance come first. The Dessa AI voice model is a different kind of bet: that the emotional layer of a conversation is worth modelling directly. Our read at StartupIndiaX is that this matters most where speech is messy. That’s where Indian voice products live, which is why Navana tracks 45 dialects.
Block’s materials we reviewed say nothing about Indian languages or accents. For any founder who gets access, that’s the first question to ask.
Here’s the short version for anyone building in voice:
- Watch the speech-first idea, not the demo polish. Block hasn’t published benchmarks, so treat the capability claims as unproven for now.
- Don’t plan a product around Dessa yet. It’s a research preview with a waitlist, and there’s no announced price or release date.
- Test distress, interruptions and impersonation in your own voice agents before customers do. A human-sounding voice raises the stakes on all three.
Building in voice AI? Tell us in the comments what you’d want a speech-first model to handle in Indian languages, and keep exploring AI and DeepTech coverage on StartupIndiaX.
FAQs
What is the Dessa AI voice model?
Dessa is a full-duplex speech-to-speech model from Block’s research lab. It listens and speaks at the same time, and Block says it was trained from scratch on human conversation, with no text and no pretrained models.
Who built Dessa and when was it announced?
A small research lab inside Block, led by Ragavan Thurairatnam, built it. The lab reports to Jack Dorsey. Block introduced Dessa as a research preview on October 23, 2026.
How is Dessa different from typical voice assistants?
Most voice assistants convert speech to text, generate a reply with a language model, then turn it back into audio. Dessa works on audio directly, which Block says preserves emotion, tone and timing that a transcript loses.
Can I use Dessa today?
Not yet. Block has opened a waitlist at block.xyz/dessa and is giving a small group of AI safety researchers early access. It hasn’t announced a broader release date, pricing or model weights.
Why does Block want AI safety researchers to test Dessa?
Block says a model trained only on speech is new territory and it wants outside testers to find risks it hasn’t anticipated. A convincing human-like voice also raises concerns about misuse, such as impersonation.
Does Dessa support Indian languages?
Block’s launch materials we reviewed don’t mention Indian languages or accents. Indian founders should treat language support as an open question until Block or independent testers publish results.
