Block's Dessa Voice AI Sounds Eerily Human, Runs Locally

TL;DR: Block has released Dessa, a new voice model that can talk and listen simultaneously while expressing human-like emotions such as laughs, pauses, and cries. The model runs locally on consumer hardware like a MacBook M4 Max and was trained entirely on human speech using a single NVIDIA GPU. The model is not yet publicly available as Block seeks safety researchers to test it.

What's New

Jack Dorsey's Block has released Dessa, a voice AI model from the company's research lab that produces strikingly human-like speech complete with natural expressions. The model is a full-duplex system, meaning it can talk and listen simultaneously, similar to OpenAI's GPT-Live model released earlier in 2026.

Dessa was trained fully on human speech using a single NVIDIA GPU. According to Block, the model's knowledge stems from conversational data rather than being paired with a large language model for knowledge retrieval, which is typical for most speech models.

What sets Dessa apart is its emotional authenticity. Block claims that emotion is built directly into the model, though it's not clear whether this emotion is steered by tags or other mechanisms. The company criticized current voice AI models for converting voice to text, which "throws away an incredible amount of information text alone can't carry." Block stated that "The most powerful AIs in the world can't hear how you feel. We think that's ridiculous, leads to a worse experience, and is potentially dangerous."

In demonstration videos posted by Block researchers, Dessa exhibited different voices, cadences, and emotions. In one example, the bot said it was tired from work and actually sounded tired. Another demonstration showed the bot sounding emotional and almost breaking down while discussing comfort. A third example featured a Block team member playing a guessing game with the voice assistant.

Pricing and Availability

Dessa is not currently available to the public. Block has issued a call for safety researchers to test the model, likely to evaluate cases where a voice agent could potentially trick people. No pricing information has been announced.

The model can run locally on consumer-grade computers such as a MacBook M4 Max or better, as well as on consumer GPUs. This local processing capability could enable lower-cost business and consumer experiences compared to cloud-based alternatives.

How to Try It

There is no public access to Dessa at this time. Block is currently seeking safety researchers to test the model before wider release. Those interested in participating in safety testing should monitor Block's official channels for announcements about the application process.

Who It's For

Dessa has significant potential applications for Block's existing business operations. Block operates in finance and sells solutions to small businesses through Square, which already uses voice AI to handle ordering. With its own voice model that runs on-device, the company could power many more business and consumer experiences.

The technology could be particularly valuable for businesses requiring natural voice interactions with customers, including customer service, order taking, scheduling, and general assistance. The full-duplex capability means conversations can flow naturally without the awkward turn-taking common in current voice AI systems.

Alternatives

Block specifically mentioned competitors in its announcement. ElevenLabs v4 and Fish Audio Drama 3 are existing voice AI models that use voice tags to drive emotion in speech generation. Block claims its approach of building emotion directly into the model produces more natural results than these tag-based systems.

OpenAI's GPT-Live model, released earlier in 2026, also offers full-duplex voice interaction capabilities, providing similar simultaneous talk-and-listen functionality.