Block Unveils Dessa, A Voice AI Trained On Human Conversation
Block says Dessa can listen and speak simultaneously, convey emotion and run locally, but access is currently limited to safety researchers.
Block has unveiled Dessa, an AI voice model it says was trained entirely on human conversation rather than written text. The company says the model can listen and speak at the same time, respond with natural pauses and emotion, and run locally on a powerful consumer computer. For now, Dessa is not publicly available.
Table of Contents
How Dessa works
Dessa is a full-duplex speech-to-speech model, meaning it can listen while it talks. According to The Verge, Block says everything the model knows came from human speech. TechCrunch reports that Block trained it using a single NVIDIA GPU rack, and that the company says Dessa draws its knowledge from conversational data rather than a separate large language model used for retrieval.
Block argues that voice systems built around converting speech to text lose information about how a person feels. Its researchers want Dessa to handle conversational timing and vocal cues more directly. The company says emotion is built into the model, though TechCrunch notes that it is unclear whether any tags are used to steer its delivery.
What the demonstrations show
Block and its researchers shared recordings in which Dessa uses different voices, cadences and emotional tones. TechCrunch describes one exchange where the model sounds tired while saying it is tired from work, as well as a guessing game with a Block team member. The Verge highlights another conversation in which Dessa responds to news of a surprise trip to Paris with what sounds like tearful gratitude.
The examples are not uniformly seamless. The Verge says a conversation shared by Dessa lead Ragavan Thurairatnam has a mostly natural back-and-forth, but the model initially appears to mistake the speaker for someone named Ronaldo. The outlet also points to an exchange about Christopher Nolan’s films in which Dessa veers into a discussion of Bryan Singer’s Superman films. These demonstrations illustrate the model’s range, but they do not establish how reliably it performs across conversations.
Access and the reason for building it
Block says Dessa can run locally on a MacBook Pro M4 Max or better, according to The Verge. TechCrunch reports that there is no public access at present. Instead, the Dessa team says a small group of AI safety researchers will receive early access.
The team lists curiosity, better understanding between people and AI, and safety as reasons for the project. It argues that voice AI should be able to recognize cues such as pain in someone’s voice rather than responding only to the words spoken. TechCrunch also notes that Block’s Square business already uses voice AI for ordering, and suggests a locally running model could support more business and consumer experiences at lower cost. Block has not announced a public release or a specific Dessa-powered product.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.