Block unveils Dessa, a full-duplex voice model it says runs locally

Block said Dessa was trained entirely on human speech and runs on a single NVIDIA GPU. Unlike many voice systems, it does not pair a separate language model with text in the speech pipeline, and Block says its knowledge comes from the conversational data it was trained on. Because it is full-duplex, it can talk and listen at the same time, similar to the GPT-Live model OpenAI released earlier this year.

Block argues that current voice AI loses emotional information when it converts speech to text. The company says Dessa builds emotion into the model itself, and it is unclear whether any voice tags steer that output. Block's researchers posted demos in which the agent changed voice, cadence and emotion during conversations.

Block says Dessa can run locally on a consumer machine such as a MacBook M4 Max or better. There is no public access yet, and Block has asked safety researchers to test the model, including cases where a voice agent could mislead the person on the other end.

If Dessa becomes available, developers could build voice agents that run on-device rather than depending on a hosted speech stack. Block has not published access terms or pricing in the article.

Developers should plan for the possibility that voice agents can convey and respond to emotion, and treat safety testing around impersonation or deception as part of the build.