The Paradigm Shift in Voice AI: How Dessa Reimagines Interaction Using Only Speech

Block’s AI research lab has pulled back the curtain on Dessa, a fully duplex conversational model that represents a radical departure in how machines learn to understand and generate human speech.

A “Voice-Only” Approach from the Ground Up

Diverging from mainstream methods, Dessa was trained exclusively on raw recordings of natural human conversations. It utilized no textual data or weights from pre-existing models throughout its development process.

This pure-audio approach allows Dessa to capture the nuances of spoken language—rhythm, pauses, emotional inflection—that are often lost in transcription. Project lead Ragavan Thurairatnam described the achievement as “a GPT-3 moment for voice.”

Remarkable Efficiency and Hardware Accessibility

Despite its sophisticated capabilities, Dessa’s computational footprint is surprisingly light. The entire training was completed using just one NVIDIA GB200 NVL72 system.

Even more striking is its deployment efficiency: the optimized model can run locally on a MacBook Pro with an M4 Max chip (or equivalent hardware) at speeds faster than real-time audio processing. This opens a path toward high-performance, low-latency voice assistants on personal devices.

  • Full-Duplex Interaction: Capable of listening and responding simultaneously for fluid conversation.
  • On-Device Operation: Local processing enhances user privacy and reduces cloud dependency.
  • High Performance: Achieves faster-than-real-time inference on consumer hardware.

Current Status and What Comes Next

Dessa is not yet publicly available. The research team has released several high-quality demo recordings showcasing its conversational abilities. Early access is being granted selectively to a small group of AI safety researchers for in-depth evaluation.

This release signals a potential shift in how voice AI systems are built. As the model matures, it could power next-generation applications in customer service, personal assistants, accessibility tools, and entirely new forms of human-computer interaction.