Block's Dessa bets on voice AI trained only on human speech
Jack Dorsey's Block has built Dessa, a voice model that never saw text. Why speech-native AI could matter most in multilingual, voice-first India.
The News
Block, the payments company led by Jack Dorsey, has unveiled Dessa, an experimental voice model that its team says was built from the ground up using only recordings of people talking. Rather than converting speech into text, reasoning over it and then reading an answer aloud, Dessa is described by its makers as a "true speech-to-speech model", with everything it knows learned from human speech.
To show it off, Block published sample conversations, several shared by Dessa lead Ragavan Thurairatnam. The clips are mostly fluid but occasionally odd: in one, the model opens by assuming the speaker is someone called "Ronaldo". Other demos include a voice guessing game, a chat about Christopher Nolan's films that drifts into Bryan Singer's Superman films, and a side-by-side exchange with OpenAI's GPT-Live-1 voice model, in which Dessa sounds the more natural of the two.
The team says Dessa can run locally on a MacBook Pro M4 Max or a more powerful machine. It has not been released publicly. Instead, a small group of AI safety researchers will get early access first.
Why It Matters
Most voice assistants today are pipelines: one system transcribes, a language model writes a reply, and a third system speaks it. That approach discards tone, hesitation and emotion, and it struggles with the rhythm of real conversation. Block's team frames Dessa around three goals - curiosity, better understanding between people and machines, and safety - arguing that an assistant deaf to distress in a caller's voice is a genuine risk.
The pitch echoes May 2024, when OpenAI's GPT-4o demo moved the industry's attention from typed chat to real-time voice. Dessa takes the idea further by dropping text from training altogether, a much harder route that the team itself concedes is difficult. It is also an unusual bet for a fintech group, hinting that Block sees voice as a future customer interface.
The local-hardware claim matters too. A model that runs on a high-end laptop, rather than only in a data centre, points towards voice agents that can operate without sending every utterance to the cloud.
Indian Angle
India is arguably the market where speech-first AI has the most to prove. Hundreds of millions of users are more comfortable speaking than typing, and everyday conversation routinely mixes Hindi, English and regional languages. Text-based pipelines often stumble on such code-mixed speech, so a model that learns directly from audio is a direction worth watching for local builders such as Sarvam AI and Gnani.ai, both selected under the government's IndiaAI Mission to develop foundation models, with Gnani focused on voice.
The financial sector is the obvious early adopter. Banks, NBFCs and insurers already deploy voice bots for customer service, KYC reminders and collections calls. A system that can detect anger or distress could help lenders handle vulnerable borrowers more carefully, something RBI's fair-practices expectations on recovery conduct already demand of human agents.
On-device processing also fits India's privacy direction. Under the Digital Personal Data Protection Act, 2023, voice recordings are personal data, and models that run locally could reduce how much customer audio has to leave a company's own systems.
FAQ
What is Dessa?
Dessa is an experimental voice model from Block, the company behind Square and Cash App. According to its team, it is a speech-to-speech model trained solely on human conversation, so it listens and responds in audio without converting speech to text in between.
Can the public use Dessa yet?
Not yet. Block says a small group of AI safety researchers will receive early access first. No wider release date or pricing has been announced; for now, the public can only hear the sample conversations on the Dessa website.
How does it compare with OpenAI's voice models?
Block published a conversation between Dessa and OpenAI's GPT-Live-1 voice model, in which Dessa came across as more natural. That is a single curated demo rather than an independent benchmark, so it should be treated as an early impression, not a definitive comparison of quality or capability.
What does this mean for Indian businesses using voice bots?
Nothing changes immediately, since Dessa is not commercially available. But it signals where voice AI is heading: native audio understanding, emotion awareness and local deployment. Indian banks, insurers and customer-service firms evaluating voice vendors may want to ask how well systems handle code-mixed speech and caller distress.
This story was reported by The Verge. Read the full original coverage at The Verge.