Introduction

Block's research division has unveiled Dessa, a full-duplex voice AI model that captures the nuances of human speech—laughs, pauses, and emotional shifts. Unlike typical voice assistants, this model processes audio directly, avoiding text-heavy pipelines that often strip away feeling and context.

What Happened

Jack Dorsey's Block released Dessa from its research lab, training the model fully on human speech using a single NVIDIA GPU rack. The model stands out as a full-duplex system, meaning it can listen and speak simultaneously, just like natural conversation. The company shared multiple demo clips where Dessa expresses tiredness, emotion, and adaptive cadence during interactions, proving its ability to convey genuine tonal shifts without relying on external text analysis.

Why This Matters

Current voice AI systems often convert speech to text, discard the raw audio nuances, then generate new speech—a process that can feel robotic and flat. Block argues that stripping away vocal emotion and context degrades the user experience and can even introduce safety risks. Dessa proves that emotion can be baked directly into the speech model, enabling more authentic interactions and opening the door for lower-cost, on-device AI that doesn't depend on constant cloud connectivity.

Key Takeaways

  • Local execution: Dessa runs on consumer hardware like a MacBook M4 Max, eliminating cloud dependency.
  • Full-duplex capability: The model enables real-time back-and-forth speech without lag, mimicking natural conversation.
  • Built-in emotion: Vocal nuance and tone are encoded in the model architecture, not stitched together via text tags.
  • Safety-first approach: Block is opening the model to researchers, with no public release yet.
  • Business impact: The technology could lower costs for companies already using voice AI, such as Square's ordering systems.

Conclusion

Block's Dessa marks a significant shift in how voice AI can be built and experienced. By putting human-like expression and local execution at the forefront, the company challenges the status quo of text-to-speech pipelines. As the line between human and machine voice blurs, Dessa offers a glimpse into a future where AI conversations feel truly natural, secure, and widely accessible.