Block releases Dessa, a full-duplex voice model trained entirely on speech that runs locally

A speech model built without text

Block, the fintech company co-founded by Jack Dorsey, has released Dessa, a voice model from its research lab that runs entirely on a local machine and is designed to sound like a person rather than a talking script. According to TechCrunch, the model was trained purely on human speech using a single NVIDIA GPU, and its delivery includes laughs, pauses and even crying.

Dessa is a full-duplex system, meaning it can listen and speak at the same time, an approach OpenAI also took with its GPT-Live model released earlier this year. Where most speech systems are paired with a large language model for knowledge retrieval, Block says Dessa's knowledge comes from the conversational data it was trained on. Dorsey promoted the launch on X, describing Dessa as the most authentic full-duplex speech model built and saying conversations with it feel like conversations with a person.

The argument against text pipelines

The technical claim at the heart of the release is that existing voice AI loses too much by routing audio through text. In a blog post cited by TechCrunch, Block argues that transcribing speech into text discards a large amount of information that text simply cannot carry, leaving even the most capable AIs unable to perceive how a user feels. The company considers that both a worse experience and a potential danger.

That is a pointed contrast with products such as ElevenLabs v4 or Fish Audio Drama 3, which use voice tags to steer emotional output during speech generation. Block claims emotion is intrinsic to Dessa rather than injected through tags, although TechCrunch notes it is not clear whether some form of tagging is still involved behind the scenes.

Demos with shifting emotion

Block's researchers published several example conversations in which the model adopts different voices, cadences and emotional states. In one, the bot mentions being tired after work and its delivery audibly matches the claim. In another, it sounds close to breaking down while telling a person that they make it feel comfortable. A third shows a Block team member playing a game with the assistant.

Ragavan Thurairatnam, one of the researchers, framed the release as a turning point for speech comparable to what GPT-3 represented for language, and noted that the model can run on consumer GPUs.

Availability and safety testing

The model is compact enough to run on consumer hardware: Block says a MacBook with an M4 Max chip or better is sufficient. There is no public access at this stage, however. Instead, the company has put out a call for safety researchers to test Dessa, which TechCrunch suggests is aimed at scenarios where a convincing voice agent could deceive the person on the other end of a call.

Why it matters

Block is fundamentally a finance company selling services to small businesses, and its Square product already uses voice AI to handle ordering. Owning a speech model that runs on-device could let Block power far more business and consumer voice experiences at lower cost, without paying for cloud inference or third-party model access.

The release also signals a broader shift in voice AI: away from pipelines that transcribe audio, reason in text and then synthesise speech, and toward end-to-end models that operate on audio directly. If Dessa's claims hold up, that approach could produce dramatically more natural conversation. It would also produce dramatically more convincing synthetic voices, which is exactly the risk Block seems to want examined before any wider rollout.