On October 23, 2026, Block, the company behind Square and Cash App, announced Dessa, a "full-duplex" speech model from a small research lab that reports directly to Jack Dorsey. TechCrunch and The Verge covered the launch on October 23 and 24. The lab says it was founded by former members of Dessa, the Toronto deep learning startup, and has built generative speech models for over eight years.
What Block says it built
The central claim is in the training data: "We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models." Block calls it "a true speech-to-speech model". It is full-duplex, meaning it listens and talks at the same time, so it can interrupt, laugh mid-sentence and backchannel with "yeah!" or "got it!" the way people do. Block says "Inference runs faster than real time", that the model "runs locally on a MacBook Pro M4 Max or better", and that it was trained "on only one NVIDIA GB200 NVL72". Block also says Dessa learned language "from hundreds of times fewer words" than an LLM needs, and that "every fact it has" came from raw sound.
The page shows live recordings, including a conversation with OpenAI's GPT-Live-1. A disclaimer says OpenAI did not review or endorse it. TechCrunch notes that other expressive systems, such as ElevenLabs v4 and Fish Audio Drama 3, steer emotion with voice tags, and says it is not clear whether Dessa's emotion is steered in some similar way.
Why Block frames it as safety
Block argues that pipelines which convert speech to text throw away tone, sarcasm and distress, and that today's AI is "still nearly deaf to the emotion in our voices". It describes that as a safety problem: imagine being in serious pain "while a voice AI has no idea and just keeps going". Block is "giving a small group of AI safety researchers early access to find the risks we haven't anticipated". Everyone else can join a waitlist. TechCrunch reports there is no public access.
What remains uncertain. Block has not published a parameter count, a paper, an evaluation or weights. "The most authentic full-duplex speech model ever built" and the comparison to LLM data needs cannot be checked. The page does not say how a model whose knowledge comes only from audio would use tools, cite sources or correct a wrong fact. Nor does it explain how text-based safeguards would apply to a model with no text inside it.
Analysis: The engineering point is that much of today's agent safety work assumes text in the loop. Moderation classifiers, prompt-injection filters, audit logs, PII redaction and evaluation graders all read transcripts. A model that goes straight from audio to audio has no transcript unless you add one, and a transcript made after the fact cannot capture tone, which is the feature being sold. Teams considering native speech models for customer service or ordering, where Square already uses voice AI, should plan for audio-native evaluation and logging. They should also decide how to handle impersonation, since a model this expressive is easier to mistake for a person. And they should keep a structured, text-based tool layer for anything that moves money or data, even if the conversation itself never touches text.
Sources and evidence
Block announced Dessa on October 23: a full-duplex speech-to-speech model it says was trained from scratch only on recorded human conversation, on one NVIDIA GB200 NVL72, and runs faster than real time on a MacBook Pro M4 Max. There are no weights, paper or benchmarks yet, only a waitlist and safety-researcher early access, and text-free models break the transcript-based moderation, logging and evaluation most voice agents rely on.