What is new

Block, the payments company run by Jack Dorsey, has announced an AI model named Dessa that it describes as trained "from scratch on nothing but human conversation." Rather than learning primarily from text and then converting speech to text and back, Dessa was built using spoken audio as its core material. Block shared several recorded conversations between Dessa and human speakers to demonstrate what the model can do, including a guessing game played entirely by voice, a discussion about Christopher Nolan's films that veers into an unplanned tangent about Bryan Singer's Superman movies, and a conversation in which a person tells Dessa about a surprise trip to Paris, to which Dessa reportedly responds with what Block's own website describes as "teary gratitude."

One clip shows Dessa briefly mistaking the person it is talking to for someone named "Ronaldo" before settling into what Block calls a mostly natural back-and-forth. Another example pits Dessa against OpenAI's GPT-Live-1 voice model in conversation, with Block saying Dessa came across as the more natural-sounding of the two.

How it works

Most voice assistants today work as a pipeline: a speech-to-text system transcribes what you say into words, a language model (trained mostly on text) decides how to respond, and a text-to-speech system turns that response back into audio. This approach is flexible and leverages huge amounts of existing text data, but it has a cost. Converting speech to text strips out tone, timing, pauses and emotional cues, and the back-and-forth timing of a real conversation can feel stilted because the system is essentially processing written sentences rather than live spoken dialogue.

Block describes Dessa as a genuine speech-to-speech model, meaning it was trained directly on audio of human conversation rather than on transcribed text. According to Block, everything the model knows came from human speech itself. The company compares this to how babies learn language: by listening and talking long before they ever see a written word. Block's team says building a model this way is possible, but very hard, which suggests the usual shortcuts of training on enormous text corpora do not directly apply when the training signal is raw conversational audio instead.

On the practical side, Block says Dessa can run locally on a MacBook Pro M4 Max "or better," meaning the model does not necessarily require sending audio to a remote data centre to function, at least for this tier of hardware.

Why it matters

Block frames the motivation behind Dessa around three things: curiosity, better mutual understanding between humans and AI, and safety. The safety argument is the most concrete one the company offers. Block's team argues that because today's voice AI systems flatten speech into text, they are effectively deaf to the emotional content of what someone is saying. They give a stark example: a person could be in serious pain while talking to a voice assistant, and the assistant would have no way of detecting that distress because it only "hears" the words, not the tone in which they were spoken. By training on raw conversational audio rather than transcripts, Block suggests a model like Dessa is better positioned to pick up on emotional and vocal cues that a text-based pipeline would simply discard.

For practitioners, this points to a broader question in voice AI research: whether the field's reliance on text-centric language models, however powerful, is fundamentally limiting how naturally and safely these systems can interact with people by voice. Dessa is being positioned as evidence that an alternative, audio-native approach is viable, even if it required extra effort to pull off.

  • Dessa is described by Block as trained entirely on human conversational speech, not text transcripts.
  • It is built by a team led by Ragavan Thurairatnam at Block, the company founded by Jack Dorsey.
  • Block says the model can run locally on a MacBook Pro M4 Max or better.
  • Block lists three motivations: curiosity, improving human-AI understanding, and safety.
  • The stated safety concern is that text-based voice AI cannot detect emotional distress, such as someone in pain, in a speaker's voice.
  • Block is giving early access to "a small group of AI safety researchers," not the general public.

Limits and open questions

The demonstrations Block has shared are a handful of curated conversations, not a systematic evaluation. There is no benchmark data, accuracy figure, or independent testing mentioned in Block's announcement to indicate how Dessa performs across a wide range of topics, accents, or emotional situations, or how often it produces odd moments like the "Ronaldo" mix-up or unplanned tangents such as the detour into Bryan Singer's Superman films. Block also does not explain in detail what "trained from scratch on nothing but human conversation" means technically: how much audio data was used, where it came from, how the model is structured, or how its safety-related claims about detecting emotion have been tested or measured.

Because Dessa is not broadly available, outside researchers and reporters cannot yet independently verify Block's description of the model's naturalness or its emotional sensitivity. The comparison with OpenAI's GPT-Live-1, where Block says Dessa sounded more natural, is also based on Block's own selected clip rather than a controlled side-by-side test.

How to try it

Dessa is not open to the public. Block says it will give early access to a small group of AI safety researchers, without specifying numbers, a timeline, or an application process. For now, the only way to experience Dessa is through the sample conversations and audio clips Block has published on its website.

Sources