Block introduces Dessa, a new voice model that mimics the human voice

Summary

Jack Dorsey's fintech company Block has introduced Dessa, a fully bidirectional (full-duplex) voice-to-voice model trained entirely on human speech that requires no text intermediary. The model can generate natural pauses, laughter, and emotional shifts, and it runs on consumer hardware like the MacBook M4 Max. There is currently no public access or independent benchmarks; Block is inviting security researchers to investigate manipulation risks.

Why it matters

If Block's claim that a voice-centric flow outperforms the dominant text-bridge architecture is validated, it will reshape the economics of voice AI for the millions of small businesses already operating on the Square platform. From a broader perspective, locally runnable voice agents with emotional expression are raising the bar for what 'natural' interaction means. This puts pressure on OpenAI, ElevenLabs, and other competitors to justify their text-centric designs. The security dimension is also critical: voice agents capable of mimicking fatigue, distress, or persuasiveness have direct implications for large-scale fraud and social engineering.

Highlights

• A fully bidirectional, voice-to-voice architecture with no text tokens in the production loop; emotion is claimed to be intrinsic rather than label-based.
• Training required only a single NVIDIA GPU rack, while inference runs on consumer hardware like the MacBook M4 Max or higher. This circumvents the costs associated with cloud-only voice AI.
• There is no public API, published benchmarks, or third-party evaluation yet; Block is issuing an open call for red-team security testing before a wider rollout.

Critical take

All performance claims rely on Block's own selected demos and a single researcher's 'GPT-3 moment' framing. Independent benchmarks and model card details (parameter count, data scale, loss metrics) are absent. The 'intrinsic emotion' claim is expressed cautiously, even by TechCrunch, which notes the ambiguity regarding whether hidden labels are in play. With no public access and a clear commercial motivation to exaggerate capabilities, this announcement reads more like a marketing move than a verifiable technical contribution.