Block unveils Dessa, a speech-to-speech AI model trained from scratch on human conversation

Block unveils Dessa, a speech-to-speech AI model trained from scratch on human conversation

Jack Dorsey's Block has announced Dessa, an AI model that the team says was trained "from scratch on nothing but human conversation." Block calls it a "true speech-to-speech model" and says everything it knows came from human speech. The model can run locally on a MacBook Pro M4 Max "or better." Dessa lead Ragavan Thurairatnam shared sample conversations, including a voice guessing game, a tangent about Christopher Nolan and Bryan Singer films, and a side-by-side exchange with OpenAI's GPT-Live-1 voice model. The Verge said the samples range from natural to awkward. In one, the model mistakes the speaker for someone named "Ronaldo." Block says it will give a small group of AI safety researchers early access.

Most voice AI today turns speech into text, reasons over the text, and then turns the answer back into speech. Block's team argues this approach loses emotion and can't keep up with how people actually time a conversation. A model trained only on audio, with no text at all, tests whether speech alone can produce a working conversational system. Block also stands out as a payments company building its own model instead of licensing one from a frontier lab. The team pitches emotional awareness as a safety feature: a voice assistant that can't tell a user is in pain is, in its words, a safety problem.

For builders, the claim that Dessa runs on a laptop matters most. If a full speech-to-speech model runs well on consumer hardware, voice products may not need cloud calls for every turn of a conversation. That would reduce latency, cost, and privacy exposure. Investors should note that Block has released no benchmarks, pricing, or wider access plans. For now, the company has only shown that the approach can work, not that it's ready for production.

#Block #Dessa #VoiceAI #SpeechToSpeech #OnDeviceAI #AISafety