←News detail

Phoenix TechnologyUpdated Oct 25 18:24

Block launches full-duplex speech model Dessa, supports local operation

AI summary

Block has launched the full-duplex speech model Dessa, which can listen to the user's speech while speaking, and can simulate emotional expressions such as laughter, pauses, and choking up. The company says it is trained entirely on human speech data, using only a single rack of NVIDIA GPUs for training, and can run locally on MacBooks with M4 Max and higher configurations. At this stage Dessa is not open for public trial; Block is sending testing invitations to security researchers, focusing on risk scenarios in which speech agents deceive humans.

Why read this

AI score 66

Block has developed its own full-duplex speech model that supports local operation on consumer-grade Macs, offering another approach to understanding on-device speech agents by comparison with mainstream speech pipelines.

Topics

Tags

Body · AI translation

Block Launches Full-Duplex AI Voice Model Dessa: Highly Human-Like and Supports Local Running

IT Home Download the Client | Exclusive Early Access October 25, 2026, 10:41:12 | From Shandong Updated October 25, 2026, 10:55:42

IT Home, October 25 news: Block, the company owned by Twitter co-founder Jack Dorsey, has launched a new voice model called Dessa. It supports local running and can simulate emotional expressions such as laughter, pauses, and choking up, sounding highly human-like. According to the official introduction, Dessa is trained entirely on human speech data, and the training process used only a single rack of NVIDIA GPUs.

Dessa is a full-duplex model. Like GPT-Live, which OpenAI released earlier this year, it can listen to the user's voice while speaking. Mainstream voice models generally need to be paired with a large language model to retrieve knowledge; Block says Dessa's own knowledge comes from the conversational data used in its training.

Block says existing voice AI experiences are often stiff and jarring, and the root cause is that the generation pipeline first converts speech to text, and then the text drives speech synthesis. Many models, such as ElevenLabs v4 and Fish Audio Drama 3, rely on voice tags to control emotion. Dorsey's company claims that emotional expression is natively built into the model, though it is currently unclear whether it still relies on some kind of tags underneath to regulate emotion.

Block wrote in its blog: "Today's voice AI interactions are really bad. They first convert your speech into text, and a great deal of information that text cannot carry is simply lost. Even the world's strongest AI cannot hear your emotions. We think this is unreasonable—not only is the experience poor, but there are also potential risks."

The research team released several sample conversations with Dessa. The agent can switch timbre, speaking rate, and emotion. In one conversation, for example, the AI says it is very tired from work, and its vocal state simultaneously carries a sense of fatigue.

In another example, the AI talks about the sense of reassurance the other person brings it, with an emotional tone that is almost choking up. Another demo shows a Block employee playing a game with the voice assistant.

According to the official explanation, the model can run locally on consumer-grade computers, with the device threshold being an M4 Max or higher-configuration MacBook. However, at this stage Dessa is not open for public trials. Block is sending testing invitations to safety researchers, focusing on testing risk scenarios in which this type of voice agent deceives humans.

According to IT Home's understanding, Block's main business is financial services, providing solutions for small and micro enterprises, and its Square subsidiary is already using voice AI to handle ordering business. After developing this voice model that can run on-device, Block is expected to launch more voice products for merchants and ordinary users at lower cost.

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of 'Dafeng Hao,' a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."

Source

Phoenix Technology