Block releases Dessa, a local full-duplex speech model
Original title: Block’s new voice model sounds eerily human
Dessa is not publicly available, and Block has asked safety researchers to test it first.
Jack Dorsey’s Block released Dessa from its research lab, a full-duplex speech-to-speech model trained only on human speech. Block says it runs locally on one NVIDIA GPU, including a MacBook M4 Max or better, and that emotion is built in rather than driven by a text pipeline. There is no public access yet. The company invited safety researchers to test cases where a voice agent could deceive a listener.
Unconfirmed: Block claims emotion is built in, but the article says it is unclear whether tags still steer it.
- Dessa can talk and listen at the same time.
- Its knowledge comes from conversational speech data, not a paired LLM.
- Square already uses voice AI to handle ordering.
- Researchers posted the announcement on October 23, 2026.
From the original article
3 min read at techcrunch.com
Block said that the voice model can run locally on a MacBook M4 Max
By Ivan Mehta · techcrunch.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.