Overview of Dessa
Block has unveiled Dessa, a new voice model developed in the company’s research lab. This model is designed to produce humanlike expressions, including laughs, pauses, and cries. Announced on October 23, 2026, Dessa is a full-duplex speech-to-speech model, capable of talking and listening simultaneously, akin to the GPT-Live model released by OpenAI earlier this year.
Training and Features
Dessa was trained entirely on human speech using a single NVIDIA GPU rack. Block has showcased examples where the agent adopts various voices, cadences, and emotions. For instance, it can sound tired when it states it is tired or exhibit emotional responses in different scenarios. The company asserts that emotion is inherently built into the model, rather than being driven by voice tags, which are commonly used in other systems.
Technical Insights
Block has provided demonstrations that include conversational clips shared by researchers, highlighting the differences between human speech and Dessa’s output. The company contrasts Dessa with existing systems that utilize voice tags, such as ElevenLabs v4 and Fish Audio Drama 3, emphasizing that Dessa’s knowledge is derived from the conversational data it was trained on, without the need for a separate large language model.
Local Operation and Current Access
The model is designed to run locally on consumer-grade hardware, such as a MacBook M4 Max or better. However, there is currently no public access to Dessa. Block is actively seeking safety researchers to evaluate the model, particularly to assess potential risks, including scenarios where a voice agent could mislead users.
Business Applications
Block’s involvement in payments and small business services positions it well to leverage Dessa. The company notes that Square already employs voice AI for ordering, and an on-device model like Dessa could enhance business and consumer experiences while reducing costs. For the time being, Block remains the exclusive distributor of access to Dessa and is prioritizing external safety testing before any public release.
Original source: TechCrunch AI