Block Unveils Dessa, a Conversation Model It Says Learned Only From Speech
Block has announced Dessa, an AI model that it says was trained from scratch on nothing but human conversation. The Verge reports that Block describes Dessa as a true speech-to-speech model, meaning everything it knows came from human speech.
Block's official Dessa page, hosted at block.xyz, is titled "Dessa - the Conversation Model trained only on speech." The page calls Dessa "the world's first AI you can have a realistic live conversation with." Those are Block's words, and the claim is the company's to make and to substantiate.
According to Block's page and The Verge's report, Dessa can run locally on a MacBook Pro M4 Max or better. Local inference, meaning a trained model running on the user's own machine rather than on remote servers, is a deployment target that changes the usual assumptions. Most conversational systems run on remote accelerators (specialized chips in data centers) and stream audio to and from an endpoint. Whether Dessa's latency (the delay between speaking and hearing a reply), throughput and memory footprint hold up at that hardware tier is a question the announcement answers only in outline.
The Dessa team said it will give early access to the model to a small group of AI safety researchers. That is a stated plan. It is not a release, and the access program is described as limited to that group.
In February 2020, Square acquired Dessa, a Toronto-based startup that builds machine learning applications. Block's press release from that time, titled "Welcoming the Dessa Team to Square", describes the company as known for its work on deepfake detection. A 2018 profile of the company lists Ragavan Thurairatnam as a co-founder. Block's 2026 materials do not, in the material I have, tie the new model to that earlier team by name, so readers should not assume a direct line of continuity from the announcement alone.
In my view, the claim that a model's knowledge comes entirely from speech will draw the closest scrutiny from the people who evaluate models for a living. A speech-to-speech system folds speech recognition, language modeling and speech synthesis into a single network. That makes its failure modes harder to isolate with the text-based benchmarks and red-teaming (deliberate attempts to make a model misbehave) that the field relies on now. Early access for safety researchers is where that scrutiny is most likely to begin, and the results of that work will say more about the model than the announcement does.
The broader context here is that conversational AI has mostly been built on text first and voice second, with speech added as an interface layer. A model trained on speech as its primary substrate, and small enough to run on a laptop, would reverse that order. If the claims hold under independent testing, the practical consequence is a conversational model that does not depend on a network connection or a cloud account to function. That would matter for privacy-sensitive deployments, for offline use, and for researchers who want to probe a model without routing its inputs through a third party's servers.
Block has announced a speech-only model, has published a page describing it, says it runs on a current high-end MacBook Pro, and plans to open early access to a small group of safety researchers. Those are the points to check first.
