Most voice assistants work like a relay race. Your speech gets turned into text, a language model writes a reply, and that reply is turned back into speech. Block thinks that’s exactly why talking to AI often feels so stiff.
Jack Dorsey’s company has unveiled Dessa, a voice model from a small research lab inside Block. The team was founded by former members of Dessa, the Toronto deep learning startup, and reports directly to Dorsey.
What makes Dessa different is how it was trained. Block says it was built from scratch on nothing but recorded human conversation. No text, and no pretrained models. In other words, it learned to talk by listening, a bit like a child does.
It’s also full-duplex, which means it can listen and speak at the same time. Block says it knows when to cut in and when to drop in small acknowledgements like ‘yeah’ or ‘got it’, and that it can hear and express emotion, sarcasm and laughter. Those cues are usually lost when speech is flattened into text.
There are some impressive practical claims as well. According to Block, inference runs faster than real time and the model runs locally on a MacBook Pro M4 Max or better. The company says it trained Dessa on a single Nvidia GB200 NVL72 rack. TechCrunch notes that, unlike many voice systems, Dessa isn’t paired with a separate large language model; its knowledge comes from the conversation data it learned from.
You can’t try it yet. There’s no public access, and Block is giving a small group of AI safety researchers early access to look for risks the team hasn’t anticipated. TechCrunch suggests one of those concerns is likely voice agents that could trick the person on the other end of a call.
Block also has an obvious business angle. As TechCrunch points out, Square already uses voice AI to handle ordering, and an on-device voice model could power more experiences at lower cost.
Why it matters
Voice is where AI meets everyday business communication: phone lines, customer support, bookings and orders. A model that sounds natural, handles interruptions and runs on local hardware could make AI phone agents cheaper and more private. That’s highly relevant for anyone running VoIP or call centre systems. It also raises the stakes for honesty: as AI voices get harder to tell apart from people, businesses will need to be upfront when customers are speaking to a machine.
Sources: Block, TechCrunch. Written by the xRapid Media LLP team.
Selling to the IT market? xRapid Media LLP is a High Tech B2B Marketing company. We help technology companies reach decision-makers with verified B2B data, account-based campaigns, email and LinkedIn outreach and appointment setting. Call +91 9767196309 or book a free B2B marketing consultation.