Block's Dessa: a voice model it says learned to talk from human speech alone

Block, the company behind Square and Cash App, co-founded by Jack Dorsey, has shown Dessa, a voice model its research lab says was trained "from scratch on nothing but human conversation". The Verge reported the announcement on 23 October; Block's page carries no publication date.

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude. Block's page includes a demo of Dessa talking with GPT-Live-1, which The Verge identifies as OpenAI's voice model, and OpenAI is a direct Anthropic competitor.

What Block says Dessa is

• Speech in, speech out. "Audio in, audio out. Dessa is a true speech-to-speech model, and everything it knows came from human speech."
• Full-duplex. Block says "it listens and talks at the same time, like you do".
• Speed and hardware. Block says "Inference runs faster than real time" and that "it runs locally on a MacBook Pro M4 Max or better."

Block also calls Dessa "the most authentic full-duplex speech model ever built". The page gives no test or measure behind that.

How Block says it was trained

Block says: "We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models." It says Dessa "learned language from hundreds of times fewer words" than large language models, and that the team "figured out how to train Dessa on only one NVIDIA GB200 NVL72", which Nvidia describes as "an exascale computer in a single rack" rather than one chip. Block does not say how much audio it used, where it came from, or how large the model is.

The team describes itself, on a page signed by Dessa lead Ragavan Thurairatnam, as "a (very) small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup", reporting directly to Dorsey.

The "first" claim, and its scope

Block writes: "As far as we know, Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech." Block hedges it and scopes it to live conversation; earlier work sits close. A 2022 research paper introduced dGSLM, which its abstract calls the first "textless" model able to generate "audio samples of naturalistic spoken dialogues", trained on two-channel conversational audio "without any text or labels"; the abstract describes generating dialogue samples, not live conversation. Kyutai's Moshi, published in 2024 as "the first real-time full-duplex spoken large language model", says it starts "from a text language model backbone", so it is not trained from speech alone.

What Block has not published

When we read the page on 24 October:

• The Download button returned the message "Download coming soon." and the waitlist button "The waitlist opens soon."
• No paper, technical report, model card or benchmark table was linked; the demonstrations are recorded conversations.
• A search of Block's GitHub organisation returned no repository matching "Dessa", and Block's verified Hugging Face organisation listed no models. With no released code or weights, there is no licence to report.

Block says it is "giving a small group of AI safety researchers early access to find the risks we haven’t anticipated".

The GPT-Live-1 demo

Two of the clips have Dessa talking with GPT-Live-1. Block's caption on both reads: "Both voices are AI-generated. The developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa." The Verge's Jay Peters wrote that in the featured exchange "Dessa comes across as more natural", and that some of Block's showcase chats "can occasionally be a little awkward and uncanny".

How Hacker News reacted

The 23 October Hacker News thread had 114 comments when we read it, none that we found from the Dessa team. A sample:

• Impressed: "This is clearly far far superior in terms of fluidity of conversation," wrote one user, who added elsewhere in the thread: "The speech realism is very good, but the intelligence is pretty clearly 2024-level." Another: "I actually thought it sounds really good and natural and it freaked me out in a good way."
• On the hardware claims: one commenter said that if Dessa can match its demos' latency and expressiveness on local devices, it is "one heck of an accomplishment". Another pointed out that a GB200 NVL72 is a full rack, adding: "The claim of “we trained this on a single rack of gpus” is definitely much less impressive than “we trained this on a single gpu”".
• Unconvinced: "The sample on the homepage sounds absolutely awful," wrote one user. Another: "Lots of style, not much substance basically." A third said that when the speaker asks the model to talk louder, "initially the voice is louder but then goes back to the same level". Block's page says: "Ask it to speak up and it gets louder."
• Can't test it: "since there is no way to confirm ie. run it, or see it working somewhere, it can be also not real," wrote one commenter.
• Misuse: several raised scam calls: "I can see this being pretty problematic for seniors/public prone to getting scammed."

Why it matters

Block argues that converting speech to text "throws away an incredible amount of information" and frames Dessa partly as a safety step, writing that AI is "still nearly deaf to the emotion in our voices". On our reading, the training approach is the notable claim and, for now, the hardest to judge. Block has published recordings, not a paper, benchmarks, weights or a working download, so outside researchers cannot yet check how well Dessa learned from speech alone.