Jack Dorsey's Block Reveals Human-Like Voice Model Called Dessa
Jack Dorsey’s company has introduced Dessa, an expressive, full-duplex speech-to-speech model trained entirely on human conversational data that can run locally on consumer hardware.
Radwan AbdulhadiEditor and technology writer
Introduction to Dessa
Jack Dorsey’s Block has officially <a href="https://x.com/jack/status/2113697902916751661">released</a> a new voice model out of the company’s research lab that sounds strangely human, featuring natural expressions like laughs, pauses, and cries. The new speech model, called Dessa, is trained entirely on human speech using a single NVIDIA GPU.
According to the company, Dessa is a full-duplex model, meaning it can listen and talk at the same time, much like the <a href="https://techcrunch.com/2026/07/08/openai-releases-new-voice-models-for-more-natural-live-conversations/">GPT-Live model OpenAI released earlier this year</a>. While typical speech models are paired with a large language model for knowledge retrieval, Block stated that Dessa draws its knowledge directly from the conversational data it contains.
Emotions and Natural Cadence
Block noted that many existing models can sound unnatural and annoying because they rely on text pipelines to drive speech generation. Platforms such as <a href="https://techcrunch.com/2026/09/28/elevenlabs-new-v4-speech-model-supports-more-expression-control-and-90-languages/">ElevenLabs v4</a> or Fish Audio Drama 3 often use voice tags to inject emotion into generated speech. In contrast, Dorsey’s company claims that emotion is natively built into its model, though it remains unclear whether tags are steering that emotion.
In <a href="https://block.xyz/dessa">a blog post</a>, the company criticized the state of voice artificial intelligence, explaining that current systems often turn voice into text and lose massive amounts of vocal information. The company argued that powerful AIs failing to perceive how a user feels can lead to inferior or potentially dangerous user experiences.
Demonstrations and Local Performance
Researchers at Block shared multiple examples of conversations featuring Dessa with varying voices, cadences, and emotions. In one demonstrated conversation, the bot mentioned feeling tired from work and adjusted its vocal tone to sound tired. Other shared clips displayed the agent playing guessing games and expressing deep nuance.
A notable technical highlight of the release is hardware capability. Block stated that the model can run locally on consumer-grade hardware, specifically a MacBook M4 Max or better, utilizing consumer GPUs.
Safety Testing and Future Business Impact
Despite its impressive capabilities and local execution, there is no public access to Dessa at the moment. Block has instead issued a call for safety researchers to evaluate the model, specifically to test for scenarios where a voice agent might trick individuals on the receiving end.
Block traditionally operates in finance and small business solutions, with subsidiaries like Square already utilizing voice AI to handle ordering. By developing an independent voice model capable of running directly on-device, the company opens up potential avenues to power expanded business and consumer experiences at a reduced cost.
Sources
- TechCrunchBlock’s new voice model sounds eerily human
Continue chronologically
Related entity coverage
Related stories
Jack Dorsey’s Block Announces Dessa AI Voice Model
Jack Dorsey’s Block has introduced Dessa, a new speech-to-speech AI model trained entirely on human conversation that runs locally on local hardware.
Block Uses OpenAI Tech to Build AI-Powered Security
Block leverages OpenAI technology to enhance its security workflows, investigating threats faster and proposing automated patches in minutes.
OpenAI, Anthropic, and Nvidia Hit Record Lobbying Quarters
Leading artificial intelligence companies ramped up their federal lobbying spending ahead of the midterms, according to new federal disclosures.
Google rolls out improved SynthID AI content detector globally
Google has launched an improved SynthID AI content detector on a new website, supporting watermarks from major partners including OpenAI and Nvidia.
