Block unveils Dessa: the voice AI that listens and talks simultaneously
The company behind Cash App and Square has unveiled Dessa, a speech processing model trained exclusively on human audio without relying on text or pretrained models. The system can listen and speak at the same time — something no conventional virtual assistant has achieved — and processes emotions like laughter, pauses, and even crying as natural parts of conversation.
The architecture behind Dessa
What sets Dessa apart from existing voice models is not just the quality of the sound, but how the system was built. While most assistants convert speech to text, process that information through a language model, and then convert the response back to audio, Dessa does everything directly in the audio domain. There is no intermediate transcript. No bidirectional conversion. The system processes sound from start to finish.
Block claims it trained Dessa from scratch using a single Nvidia GB200 NVL72 computing system — a massive machine combining 72 H100 GPUs interconnected by NVLink. The company also stated that the model is compact enough to run locally on a MacBook Pro M4 Max or higher, without requiring cloud infrastructure.
The approach eliminates the massive information loss that occurs when human speech is converted to text — accents, irony, hesitation, and emotional intensity simply disappear. Rodion Gurev, Block's CTO, told TechCrunch: "We believe voice is the most natural way to interact with artificial intelligence, and we find it ridiculous that the most powerful AIs in the world cannot sense how you feel by your tone of voice."
What the model can do
The demonstration examples published by Block's research team show live recorded conversations without a script, where the model spontaneously reacts to surprises, makes natural pauses, laughs mid-sentence, and even expresses tiredness when asked — changing the tone and rhythm of the voice to match the narrated state.
Jack Dorsey wrote on his social networks that "talking with it feels like talking with a person" — full of emotion, nuances, pauses, and natural rhythm. The model is described as full-duplex, meaning it can process audio input and output simultaneously. If you interrupt Dessa while it is speaking, it stops immediately and processes the user's speech, without needing to press a button to change turns, as happens with Siri, Alexa, or Google Assistant.
Ragavan Thurairatnam, a Block researcher, classified the milestone as "a GPT-3 moment for voice processing" — a direct reference to the point when GPT-3 demonstrated that sufficiently large language models could perform tasks for which they were not explicitly trained. The analogy suggests that Dessa may represent a similar shift for the voice domain, but with potentially broader implications, since verbal conversation is humanity's primary form of communication.
Why this matters: security and fraud implications
The same capability that impresses enthusiasts concerns security experts. Justine Bateman, filmmaker and technology critic, questioned why artificial voices need to sound human enough to deceive people. "That's how you can trick people into lowering their guard," she wrote.
Shiv Aroor, a technology journalist, was more direct: "Fraud calls are about to hit pandemic levels." The concern is not speculative — voice cloning technology has already been used in phone scams in 2025 and 2026, including cases where victims were tricked into transferring money or revealing personal information based solely on the caller's voice. A model like Dessa, capable of generating voices indistinguishable from humans with near-zero latency, exponentially amplifies that risk.
Block acknowledged this problem. Rather than launching the model to the public, the company gave limited access to a small group of AI safety researchers, asking them to test scenarios where a voice agent could deceive people. No public model weights, benchmarks, or API have been released so far.
Dessa's legacy and the future of voice interaction
The name Dessa comes from the Canadian startup that Square acquired in 2020, which was already working on natural speech processing. The acquisition was widely seen at the time as strategic, but the product now is far more ambitious than any commercial speech processing system available.
The broader implication is that, as voice models like Dessa become more accessible and efficient, voice interaction may surpass the keyboard as the primary interface for artificial intelligence interaction — especially for tasks requiring emotional sensitivity, such as education, mental health, and customer service. Binance founder Changpeng Zhao has already suggested the model could be particularly useful in children's education, where conversational learning surpasses textual instruction.
The question is whether this new ability to understand and express emotion through voice will make AI interactions truly more human — or whether it will simply create the most convincing manipulation tool ever built.
Sources: TechCrunch, CoinDesk, LeoDex News, CryptoNews, Block official blog
✓ Independent sources cross-checked and verified before publishing
