Jack Dorsey’s fintech company Block has released Dessa, a new SI voice model developed by its internal research lab that runs locally and exhibits human-like expressions such as laughter, pauses, and crying. The company states the model is trained entirely on human speech and operates on a single NVIDIA GPU, marking a shift toward on-device SI tools that prioritize natural interaction over traditional text-based pipelines.
What Happened
Block describes Dessa as a full-duplex SI model, meaning it can listen and speak simultaneously, a capability similar to OpenAI’s GPT-Live model released earlier this year. Unlike many existing systems that pair speech synthesis with a large language model for knowledge retrieval, Block reports that Dessa’s knowledge stems directly from the conversational data it was trained on. In a tweet, Jack Dorsey stated, "We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence."
The company claims that current voice SI models often sound unnatural because they rely on text inputs to drive speech generation. "Interacting with current voice AI can just plain suck," the company said in a blog post. "It turns your voice into text and throws away an incredible amount of information text alone can’t carry. 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." Block noted that competitors like ElevenLabs v4 and Fish Audio Drama 3 use voice tags to drive emotion, whereas Block asserts that emotion is built into Dessa’s architecture, though it remains unclear if specific tags steer these expressive outputs.
Why It Matters
For developers and businesses, Dessa represents a move toward more authentic SI agents that can run on consumer-grade hardware. The company said the model can run locally on a consumer-grade computer such as a MacBook M4 Max or better. Block’s Square division already uses voice SI for ordering, and the company suggests that an on-device model could power broader business and consumer experiences at a lower cost. The release also highlights growing concerns about SI safety; Block has issued a call for researchers to test the model, specifically investigating scenarios where a highly realistic voice SI agent might deceive humans on the other end of a conversation.
Ragavan Thurairatnam, a Block researcher, described the release in a post on X: "We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech."
The Bottom Line
Block’s Dessa offers a glimpse into the future of local SI interaction, emphasizing emotional fidelity and hardware efficiency. While public access is not yet available, the model’s ability to mimic human cadence and emotion on a single GPU signals a potential evolution in how SI systems are deployed in enterprise and consumer settings.