Block Launches Dessa AI Voice Model Trained Purely on Human Speech
Trained entirely on spoken conversations rather than written text, Block's experimental system runs locally on Apple laptops while, according to Block, preserving natural emotional cues.
In short
- Block announced Dessa, a full-duplex speech-to-speech model trained strictly on recorded human speech with zero text data or pretrained LLMs.
- The model was trained on a single Nvidia GB200 NVL72 server rack and can run locally on an Apple MacBook Pro with an M4 Max chip or better.
- According to Block, unlike typical text-based voice pipelines, Dessa detects and responds to emotional nuance, laughter, sarcasm, and spontaneous interruptions in real time.
- Public access is currently unavailable while Block distributes early builds to independent AI safety researchers.
Financial technology company Block has introduced an experimental voice system that circumvents written language altogether. On October 23, 2026, the company unveiled the Dessa AI voice model, an audio-to-audio architecture trained from scratch exclusively on spoken human dialogue. Rather than transcribing audio to text, running it through a conventional large language model, and synthesizing an artificial response, Dessa processes and generates sound directly. According to announcement details shared by The Verge, the project aims to address how traditional text-bound pipelines fail to keep up with human timing and remain deaf to emotion and pain.
The initiative emerges directly from Block's internal research divisions, led on the project by Dessa team lead Ragavan Thurairatnam. By teaching an artificial intelligence model to understand natural spoken dialogue without textual intermediary layers, Block claims to have addressed long-standing limitations in real-time vocal communication.
How the Dessa AI Voice Model Works Without Text
Most modern conversational voice agents convert speech into text and rely on text in the pipeline to drive speech generation. According to reporting by ANI, Block discarded this entire structure. Dessa is a full-duplex speech-to-speech model, meaning it simultaneously listens and speaks over an active audio stream without waiting for turn-taking boundaries.
Block claims that emotion is built directly into its model without text in the pipeline driving speech generation, though TechCrunch reported it remains unclear whether emotion is steered by tags. In a company statement highlighted by TechCrunch, Block explained that standard text-first voice tools strip out vital acoustic information. When spoken audio is flattened into plain text, the system becomes oblivious to pain, crying, hesitation, or sarcasm. Dessa was designed to perceive these cues natively from audio directly, enabling spontaneous laughter mid-sentence or dynamic volume adjustments when asked to speak louder.
Block stated that training on sound required substantially fewer word instances than modern text models require. Block noted that spoken communication contains far denser contextual signals per utterance, allowing Dessa to acquire conversational flow from hundreds of times fewer words than traditional text baselines.
Hardware Efficiency and Local Execution
Block reported constrained compute requirements for Dessa's foundation. According to coverage in Notebookcheck, Block trained the entire model using a single Nvidia GB200 NVL72 rack containing 72 Blackwell GPUs.
Equally notable is Dessa's target inference footprint. The company stated that the model can run locally on consumer-grade hardware, specifically an Apple MacBook Pro configured with an M4 Max processor or higher. Block claims local inference operates faster than real time, allowing conversations without noticeable pauses.
As detailed by Notebookcheck, Block has not yet released the exact parameter count, model weights, or formal evaluation benchmarks for Dessa.
Conversational Demonstrations and Behavioral Flaws
To exhibit Dessa's capabilities, Block published a collection of audio and video demonstrations recorded with staff members. As observed by Gizmodo, the model demonstrated fluid persona adaptation and realistic acoustic mannerisms, though not without revealing operational quirks.
In one demonstration, a Block team member played a guessing game where Dessa correctly identified an everyday object on the first attempt twice in a row, yet continued offering new guesses after being told it had already won. In another recording, Dessa adopted the vocal cadence of an African American man describing exhaustion from farm labor and proudly sharing the birth of a newborn son named Darius Johnson. Additional audio samples showed Dessa laughing spontaneously, discussing Christopher Nolan movies before diverting onto Bryan Singer's superhero films, and sounding overcome with tearful gratitude when told it was being taken on a trip to Paris.
Block also showcased a direct interaction between Dessa and OpenAI's GPT-Live-1 audio model, as reported by The Verge. In that dialogue, Dessa displayed a more natural conversational rhythm with realistic pauses, contrasting with the more rigid cadence of OpenAI's system. However, in a separate conversation shared by Ragavan Thurairatnam, the model made an error by initially mistaking a human speaker for somebody named "Ronaldo".
Research Intent, Safety Oversight, and Practical Outlook
Block attributed its decision to build Dessa to three distinct motivations: technical curiosity, improving human-machine interaction, and addressing safety failures in existing voice interfaces. The engineering team argued that if an AI system cannot hear panic, pain, or distress in a user's voice, it poses a safety hazard in high-stakes environments.
Despite the demonstrations, Dessa is not available for public download or general commercial deployment. According to The Decoder, Block opened a sign-up application specifically for AI safety researchers. The company stated that external scrutiny is essential to uncover unpredicted risks.
TechCrunch pointed out that Block could eventually integrate on-device speech technology into its merchant hardware. Square already uses voice AI to handle ordering. An on-device voice model could power business and consumer experiences at a lower cost.
Questions people ask
What is Block Dessa?
Dessa is an experimental full-duplex speech-to-speech AI model developed by Block. It is trained from scratch purely on recordings of spoken human conversation, without relying on written text or pretrained large language models.
What hardware is required to run Dessa locally?
According to Block, Dessa can execute local inference on an Apple MacBook Pro equipped with an M4 Max processor or higher, running faster than real time without cloud infrastructure.
How was the Dessa model trained?
Block trained the model on a single Nvidia GB200 NVL72 server rack housing 72 Blackwell GPUs, using speech audio exclusively.
Is the Dessa voice model available to the public?
No. Block has limited early access to selected AI safety researchers via a sign-up form to identify unforeseen risks before considering any broader release.
Sources
- Jack Dorsey’s Block made an AI model that talks to you like an awkward new friendThe Verge, Oct 23, 2026
- Jack Dorsey's Block launches Dessa; AI voice model trained entirely on 'human conversation'ANI, Oct 24, 2026
- Block's new AI model Dessa was trained entirely on human speech to catch emotions text-based models missThe Decoder, Oct 24, 2026
- Block’s new voice model sounds eerily humanTechCrunch, Oct 24, 2026
- Check Out the Many Uncanny, Human-ish Voices From Block's New AI Modelgizmodo.com, Oct 24, 2026
- Jack Dorsey's new AI never read a single word, yet it can hold disturbingly human conversationsNotebookcheck, Oct 24, 2026
How this story was made: the newsroom picked it up from theverge.com, Hacker News and Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (22 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 25, 2026 at 02:14 UTC
