Block introduces Dessa, a local, full-duplex voice model
Block has released a new research-stage voice model called Dessa, capable of full-duplex, emotional speech synthesis that runs on local consumer hardware.
Conceptual illustration, not a product screenshot: Bakken & Baeck / Google DeepMind, Visualising AI. https://unsplash.com/photos/a-black-and-white-photo-of-a-wall-r0xgEG0Teu0 | Unsplash License: https://unsplash.com/license / رسم مفاهيمي مرخص وليس لقطة للمنتج.
TechCrunch reported on October 24, 2026, that Block's research lab has developed Dessa, a full-duplex speech-to-speech model. Unlike systems that rely on text-to-speech pipelines, Dessa is trained entirely on human speech data, allowing for integrated emotional nuances like pauses, laughs, and variable cadences. How emotional expression is directed remains unconfirmed here.
As a full-duplex system, the model is designed to both talk and listen simultaneously, aiming for a more natural, conversational interaction flow. The developers have indicated that the model is intended to run locally on consumer-grade hardware, such as the M4 Max chip, potentially enabling on-device applications without the latency typical of cloud-based retrieval systems.
For those testing voice-based applications, the key takeaway is the potential for local, low-latency interaction. As this remains in a research phase with limited access for safety testing, developers can begin considering how local inference changes their architecture, particularly regarding how emotional delivery is handled when text is removed from the generation pipeline.
Source: TechCrunch