Block unveils Dessa full-duplex voice AI model
Block has introduced Dessa, a full-duplex speech AI model capable of listening and talking simultaneously with emotional nuance and local hardware execution.
Stock photo for illustration only, not from the actual event
- Block launches Dessa, a new full-duplex speech-to-speech AI model.
- The model relies on conversational data instead of an external LLM.
- Dessa can run locally on consumer hardware like a MacBook M4 Max.
- Public access is currently restricted as Block enlists safety researchers.
Block has introduced Dessa, a full-duplex speech model designed to listen and talk at the same time, mirroring capabilities seen in OpenAI's GPT-Live model released earlier in the year. Unlike traditional speech models that typically pair with a large language model for knowledge retrieval, Block states that Dessa derives its knowledge directly from its conversational training data.
In a published blog post, the company criticized existing voice AI systems for converting speech into text and losing critical emotional information along the way. While models such as ElevenLabs v4 or Fish Audio Drama 3 utilize voice tags to drive emotional generation, Block claims that emotion, pauses, and cadence are built directly into Dessa without clear reliance on external tags.
Stock photo for illustration only, not from the actual event
Moving from text-reliant pipelines to pure speech-to-speech architectures represents a major shift in conversational AI, aiming to eliminate robotic cadences. For Block, integrating this on-device capability could significantly enhance consumer and business applications—such as Square's automated ordering systems—while cutting operational costs.
Block's researchers demonstrated multiple conversation samples showing Dessa adapting its voice, cadence, and emotional state. In one instance, the bot expressed fatigue from work with a matching tired tone, while another demonstrated emotional delivery nearly breaking down. A separate demonstration featured a Block team member playing a guessing game with the AI assistant, which successfully guessed answers promptly.
Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can’t carry. The most powerful AIs in the world can’t hear how you feel.
Block
Furthermore, Block highlights that Dessa is trained entirely on speech and can run locally on consumer-grade hardware like a MacBook M4 Max or better. Public access is currently unavailable, as Block has initiated a call for safety researchers to evaluate the model and test potential risks where voice agents could be misused to deceive individuals.
Source: TechCrunch
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment