What Block demonstrated
Block has shown Dessa, a research voice AI model whose synthetic speech includes pauses, laughter and other expressive cues. As TechCrunch reports, the system is designed for full-duplex conversation: it can listen while speaking, making interruptions and timing feel more natural.
Why speech-to-speech matters
Block says Dessa works directly with speech rather than depending on a text transcription at the center of each exchange. The company argues that conventional speech-to-text pipelines can lose tone and timing. This design may explain why its demonstrations sound less like rigid turn-taking, though the clips are company examples rather than independent performance tests.
What remains unverified
Block says it trained Dessa on a single NVIDIA GPU rack and that it can run locally on a MacBook M4 Max or better. Those are company claims. Dessa is not publicly available, so outside users cannot yet test the model broadly. A voice that sounds tired or amused is not evidence that a model feels emotion or has humanlike cognition.
Where this could lead
More fluid voice interaction could make assistants easier to use, while highly convincing speech raises questions about impersonation and trust. Block has invited safety researchers to evaluate Dessa. It has not announced the model as a Square product. Watch the Daily AI Chat Short or listen to the full podcast episode. Source: Ivan Mehta, TechCrunch, October 24, 2026.
