Block's Dessa voice AI adds full-duplex emotion to challenge assistants
Block has released Dessa, a full-duplex voice AI model trained solely on human speech that runs locally on consumer hardware, processing audio directly to preserve emotional nuance and address critical safety gaps in current voice technology.

Block has released Dessa, a full-duplex voice AI model that runs locally on consumer hardware and demonstrates human-like conversational behaviors. The model, trained from scratch only on human conversation data, processes audio directly without converting speech to text, aiming to eliminate the emotional blind spots of current assistants.
Dessa is a full-duplex speech-to-speech model capable of talking and listening simultaneously, similar to OpenAI's GPT-Live model. It can run on a single Nvidia GPU rack or on consumer-grade computers like a MacBook M4 Max, where it processes inference faster than real time. The model was developed by a small AI research team at Block led by Ragavan Thurairatnam, founded by former employees of a company also named Dessa. In demonstrations, Dessa exhibited behaviors like laughing mid-sentence, adjusting its volume based on requests, expressing emotional responses to surprises, and making contextual errors such as misnaming a user.
Why current voice AI falls short
Block has been critical of existing voice AI, describing it as ridiculous and potentially dangerous. The company states current systems can just plain suck because they convert voice to text, losing crucial vocal information and emotional nuance in the process. This limitation creates safety concerns, particularly in scenarios where an AI cannot detect human distress. Jack Dorsey introduced Dessa on Friday, October 23, framing its development as a direct response to these shortcomings.
Dessa's technical breakthroughs
Dessa's architecture bypasses text conversion entirely, processing raw audio to preserve vocal subtleties. Emotion is built into the model, allowing it to express different voices, cadences, and emotions, such as sounding tired or emotional. The model can handle multiple voices and emotions in a single conversation. Block's website includes live recordings of real people interacting with Dessa without human steering, showcasing these capabilities. Its knowledge is derived exclusively from the conversational data on which it was trained, with no books, web pages, or pretrained models used.
Safety and industry implications
The development of Dessa directly addresses critical safety gaps in voice AI, particularly regarding emotion detection and human distress scenarios. Block is currently seeking AI safety researchers to test the model, as it is not yet publicly available. Wider release is planned following this early access phase. The launch is highly relevant to current AI safety debates, as researchers urgently evaluate voice AI's emotional blind spots. Within Block's own ecosystem, its Square division already uses voice AI to handle ordering, and Dessa is reportedly the first voice AI that restaurant owners cannot identify as machine-based.
Block, a company in finance that sells solutions to small businesses, reported $2.42 billion in bitcoin revenue for a single quarter. In a related statement, Jack Dorsey commented that bitcoin fails through irrelevance if it remains only a store of value. Dessa's early access is currently limited to AI safety researchers, with a wider release planned.





