Jack Dorsey's Block Unveils Local Human-Like Voice AI Model Named Dessa

Jack Dorsey's company Block has released a new voice model called Dessa from its research lab. The model runs locally on a consumer-grade computer such as a MacBook M4 Max or better using a single NVIDIA GPU. According to Indian Express - Technology, the system inference runs faster than real time with no latency. Dessa is a full-duplex speech-to-speech model, meaning it can talk and listen at the same time without any transcription layer in between. Block states that the model is trained fully and directly on human speech data without any text or pretrained models involved. Unlike other speech systems that rely on a separate large language model for knowledge retrieval or use text and voice tags to drive emotion, Dessa builds emotion directly into the model. The company states that the tool features natural human expressions like laughs, pauses, cries, and a natural cadence. In shared demo examples, the bot sounds tired after work or gasps like a real person while holding a live conversation. Block claims that current voice AI solutions often feel annoying because they turn voice into text and throw away rich emotional information. Public access to Dessa is not available at the moment. Block has put out a call for safety researchers to test the model behind a waitlist to evaluate possible risks such as voice agents tricking people. The company currently has no parameter count, API, price, weights, or published evaluation details. Block operates in finance and sells solutions to small businesses, with Square already using voice AI to handle ordering. Running an on-device model can help power more business and consumer experiences at a lower cost.

Context

Voice assistant technology has traditionally relied on converting spoken words into text before processing them, which loses the subtle emotional cues of human speech. On-device models running locally on personal laptops mark a shift away from heavy cloud servers for audio processing. Companies are increasingly exploring how voice interfaces can handle natural, two-way conversations without awkward delays.