Block Unveils Dessa: Full-Duplex Voice Model Runs Locally on Consumer GPUs

Block's research lab has released Dessa, a full-duplex speech-to-speech model that runs locally on consumer-grade hardware such as an NVIDIA GPU or a MacBook M4 Max. Unlike conventional pipelines that convert speech to text, process it with an LLM, then synthesize speech, Dessa was trained end-to-end on raw human conversation. The result is a model that laughs, hesitates, cries, and matches cadence in real time, all without explicit emotion tags.

Why it matters for GPU / AI infrastructure

Dessa demonstrates that high-fidelity, emotionally aware voice interaction no longer requires massive data-center clusters. By fitting a full-duplex model onto a single consumer GPU, Block shifts the cost and latency profile of voice AI toward edge deployment. For AiGpu customers, this signals a growing class of workloads — real-time speech-to-speech — that can be served efficiently on mid-range accelerators, reducing both capex and round-trip latency for applications in finance, retail, and customer support.

Block emphasizes safety and has opened a call for researchers to stress-test the model before any public release. The company already uses voice AI for Square's ordering systems; an on-device model like Dessa could expand those capabilities to offline or privacy-sensitive environments while cutting cloud inference costs.