Block

Artificial Intelligence · Voice AI · Fintech

Launch

Block unveils Dessa, a full-duplex voice AI trained purely on speech

It learned language from a single Nvidia chip, suggesting voice AI could need far less computing power than today's text-based models.

• Block, the fintech company led by Jack Dorsey, unveiled Dessa, a full-duplex speech-to-speech AI model trained entirely on human speech, on October 23, 2026.
• A small AI research lab inside Block, led by Dessa lead Ragavan Thurairatnam, built the model to capture tone, emotion and timing that text-based voice assistants typically lose.
• Dessa processes audio directly with no transcription layer, letting it laugh mid-sentence, interrupt naturally and adjust its volume when asked, behaviors Block says nobody explicitly taught it.
• Block says Dessa was trained on a single Nvidia GB200 NVL72 system and learned language from "hundreds of times fewer words" than text-based LLMs typically require.
• The model runs locally on a MacBook Pro M4 Max or better without an internet connection, and is currently a research preview behind a waitlist with no published parameter count, API, pricing or benchmark scores.
• Mainstream voice assistants like those from OpenAI and Google first convert speech into text before replying, stripping out emotional cues; Dessa skips that step by learning directly from raw audio.
• Block is granting early access to a small group of AI safety researchers through its waitlist before any wider rollout or commercial release.
• By training on raw audio instead of text, Block is testing whether speech itself can be a more data-efficient and emotionally expressive foundation for AI than the transcription-based pipelines most assistants rely on.