Block Unveils Dessa, a Lifelike Voice AI That Runs Locally
Block unveils Dessa, a local full-duplex voice AI trained only on human speech, with lifelike emotion, consumer GPU support and no public access for now.
Summary
Jack Dorsey’s Block unveiled Dessa from its research lab on October 23, 2026, describing an expressive speech-to-speech model trained entirely on human speech using a single NVIDIA GPU. The full-duplex system listens and talks simultaneously, matching the core interaction of OpenAI’s GPT-Live, released earlier in 2026. Unlike speech products paired with a large language model for retrieval, Dessa draws its knowledge from conversational training data. Block says removing text from the pipeline retains cues that enable laughter, pauses, crying and natural cadence. ElevenLabs v4 and Fish Audio Drama 3 use voice tags to shape emotion, but Block has not disclosed whether tags steer Dessa despite claiming emotion is built in.
Demonstrations showed multiple voices, cadences and emotions: Dessa sounded tired while discussing work, nearly broke down while talking about being made comfortable, and played a guessing game with Block team member Cara. The model runs locally on consumer GPUs, including a MacBook M4 Max or better, potentially reducing costs and broadening Block’s finance, small business and consumer products. Square already uses voice AI for ordering. As of October 24, 2026, Dessa has no public access, and Block is recruiting safety researchers to test it, potentially including whether its lifelike voice can deceive people.
Positives
- A single NVIDIA GPU trained Dessa entirely on human speech, while consumer hardware can run the finished model locally.
- Full-duplex operation lets Dessa listen and speak simultaneously, enabling more natural conversational timing.
- Demonstrations showed tiredness, emotional distress, varied cadences and guessing-game interactions without a text-driven speech pipeline.
- MacBook M4 Max support could lower deployment costs and expand Block’s finance, small business and consumer voice products.
- Square’s existing voice ordering system gives Block an established commercial use for its on-device model.
Risks & concerns
- No public access was available as of October 24, 2026, preventing independent evaluation of Dessa’s performance claims.
- Block has not clarified whether hidden voice tags steer emotion despite claiming emotional expression is built into Dessa.
- Dessa relies on conversational training data rather than a separate large language model retrieval layer, leaving its practical knowledge range unspecified.
- Block’s safety testing may examine whether Dessa’s lifelike speech can deceive people, highlighting a potential manipulation risk.