Block, the company co-founded by Jack Dorsey, has unveiled Dessa, a groundbreaking full-duplex AI voice model designed to deliver incredibly human-like vocal interactions. What sets Dessa apart is its ability to run locally on consumer hardware, specifically Apple MacBooks equipped with M4 Max chips or higher.
True Emotional Expression in AI Voice
Unlike many existing voice AI models that often sound stiff and unnatural, Dessa is trained entirely on human speech data. Block emphasizes that the model can naturally incorporate emotional nuances such as laughter, pauses, and even vocal catches, making conversations feel remarkably more authentic. This contrasts with traditional approaches where AI voice generation often involves a text-to-speech pipeline, which can strip away subtle emotional cues. Many current models rely on voice tags to control emotion, but Block claims Dessa’s emotional expressiveness is an inherent part of its design, built directly into the model.
“The current voice AI interaction experience is terrible. It first converts your speech to text, losing a lot of information that text can’t carry. Even the world’s strongest AI can’t hear your emotions. We believe this is unreasonable, not only is the experience poor, but there are also potential risks.”
– Block
Block highlights the limitations of current voice AI, stating that the text conversion process inherently loses a significant amount of information. Even advanced AI struggles to detect human emotion, leading to frustrating and potentially risky interactions. Dessa aims to bridge this gap by processing voice directly, preserving the rich emotional data present in human speech.
Full-Duplex Capability and Local Operation
Dessa operates as a full-duplex model, similar to OpenAI’s GPT-Live, meaning it can listen and respond simultaneously. This enables a more natural conversational flow, where the AI can react in real-time to the user’s input without noticeable delays. The ability to run locally on MacBooks is a significant achievement, reducing reliance on cloud infrastructure and potentially enhancing privacy and reducing latency.
Demonstrations Showcase Versatility
To illustrate Dessa’s capabilities, Block has released several demonstration dialogues. These examples showcase the AI’s ability to dynamically adjust its tone, speed, and emotional expression. In one instance, the AI conveys a sense of tiredness in its voice, while another demonstrates a comforting tone that borders on emotional. A further demonstration shows the AI engaging in a game with a Block employee, highlighting its interactive potential.
Future Applications and Testing
While Dessa is not yet available for public testing, Block is inviting security researchers to evaluate its potential risks, particularly concerning scenarios where voice AI might be used to deceive humans. Block, primarily known for its financial services and solutions for small businesses through Square, sees significant potential for Dessa. The company is already using voice AI for order processing within Square. Developing a self-researred, locally runnable voice model like Dessa could enable Block to deploy a wider range of voice-powered products for both merchants and consumers at a lower cost.
The development of Dessa signifies a leap forward in creating more engaging and natural voice AI interactions, moving beyond mere information retrieval to genuine conversational partners.
