Learned only from conversations: Dorsey's Block unveils voice AI Dessa – fraud warnings
Block has unveiled a voice AI trained only on human speech that listens and talks at the same time. For now, only safety researchers are allowed to test it.
Archive image · Photo: JD Lasica / Wikimedia Commons, CC BY 2.0 (cropped and resized)
Payments company Block, run by Twitter co-founder Jack Dorsey, unveiled a research preview of a speech model called Dessa at the end of the week. According to the company, it was trained from scratch on human conversation alone – with no text and no pretrained models. Dessa processes audio directly, listens while it talks, drops in short acknowledgements such as “yeah!”, knows when to interrupt and is said to recognise and express emotions such as sarcasm, laughter and sighs.
Block says Dessa was trained on a single Nvidia GB200 NVL72 system and runs faster than real time locally on a MacBook Pro with an M4 Max chip or better. Block did not disclose the model's size, benchmark results or a release date. It was developed by a small AI research lab that includes former staff of Toronto start-up Dessa, which Square acquired in 2020. Dorsey wrote on X that talking with Dessa feels like talking with a person.
Initially, only a small group of AI safety researchers is getting access to find risks the team has not anticipated. Reactions ranged from enthusiasm to concern: several voices on X warned, according to CoinDesk, that such a human-sounding voice could help phone scammers. Block already uses voice AI for ordering in its Square payments service, TechCrunch reports.