Block releases speech AI trained on 100% human audio

Block, the financial technology firm led by Jack Dorsey, announced the release of Dessa, an artificial intelligence model trained exclusively on human conversation. The company stated that the system was built "from scratch on nothing but human conversation," meaning no written text was used during the training process.

Ragavan Thurairatnam, who leads the project, shared sample audio files showing the model interacting with users. In one recording, the model initially mistakes the speaker for a person named Ronaldo, but the dialogue quickly settles into a natural flow. The Dessa website features additional clips, including one where the model discusses Christopher Nolan films before tangentially referencing Bryan Singer’s Superman movies. Another example shows the model responding to news of a surprise trip to Paris with what the site describes as teary gratitude. A separate clip pairs Dessa with OpenAI’s GPT-Live-1 voice model, where the speech appears more fluid.

The development team describes Dessa as a true speech-to-speech model. They claim the system can run locally on a MacBook Pro M4 Max or better hardware, suggesting it does not require constant cloud connectivity for basic operation. This local capability distinguishes it from many cloud-dependent voice assistants.

The team cited three primary motivations for the project: curiosity, improved human-AI understanding, and safety. They compared the training method to human development, noting that babies learn to talk without seeing written words. The developers stated, "It can. It's just very hard," referring to the challenge of training a model on audio alone.

Block argued that current voice assistants reduce the value of speech by converting it to text. The team said this process flattens the nuances of conversation and fails to match human timing. They positioned speech as the most natural interface for intelligence and suggested that text-based intermediaries create delays and lose emotional context.

The safety argument focuses on emotional perception. The team noted that AI systems are gaining power while remaining deaf to the emotion in human voices. They described a scenario where a user is in serious pain but a standard voice AI continues its task, unaware of the distress. The team called Dessa a step toward fixing this gap.

Block plans to provide early access to a small group of AI safety researchers. The announcement did not include a timeline for a wider public release or specific pricing details.