EDITION OF OCTOBER 25, 2026
Dessa: Block unveils a local voice AI, still without public access
Block has unveiled Dessa, a voice model designed to listen and speak simultaneously, with local operation announced. Demonstrations highlight expressive variations, but the lack of public access and independent evaluation in the available information limits conclusions about its performance.
Legal disclosure: this content was produced by an artificial intelligence system. Editorial transparency
With Dessa, Block highlights expressive voice conversation and operation on a personal computer. The presentation raises a practical question: how do you evaluate an assistant whose voice sounds natural? The available information describes the model's ambitions, but its performance and control still need to be examined.
Block has unveiled Dessa, a voice model developed by its research lab, TechCrunch reports. The company describes it as a system capable of running locally and producing expressive speech, with laughter, pauses or vocal sounds resembling crying. These features are presented by Block and illustrated in its demonstrations; they do not constitute an independent measure of how natural the exchanges are.
The technical choice highlighted is full-duplex operation: according to the presentation, Dessa can speak and listen simultaneously. Block also says its training relies entirely on human speech. The company explains that the model's knowledge comes from its conversational data, whereas voice systems are generally paired with a large language model to access knowledge, according to TechCrunch.
This approach puts the way speech is delivered at the center of the offering. Block believes that converting speech to text loses some of the information conveyed by intonation and rhythm. Its researchers have published several conversations featuring different voices, speaking rates and expressions. These examples show what the company wants to highlight, without establishing how consistently these behaviors occur across different situations.
Control over this expressiveness also remains to be clarified. TechCrunch notes that it is unclear whether vocal emotions are controlled through tags. A voice that sounds tired or emotional is therefore insufficient to demonstrate that the system correctly interprets the state of the person it is speaking with. A subsequent evaluation would need to distinguish the quality of the audio delivery from the relevance of the responses.
Block says the model can run on a consumer computer such as a MacBook M4 Max or better. This possibility could interest applications seeking to process exchanges on the device, but the available information does not establish the hardware requirements or the performance achieved in this use case.
However, Dessa was not publicly accessible at the time described by TechCrunch. Block has invited security researchers to test the model. At this stage, the presentation therefore sets out a research direction and selected demonstrations. It leaves open questions about reliability, control over expressiveness and behavior when faced with less predictable conversation partners or requests.