Local AI Speech Model: Block Releases Dessa

Block, the technology company led by Jack Dorsey, announced on October 23 that it has released an artificial intelligence model named Dessa. The system is designed to process and generate speech directly without converting audio into text first. Block states that the model was trained exclusively on human conversation, a method the company says allows it to capture nuances that text-based systems often miss.

The first key detail is the hardware requirement. Dessa runs locally on a MacBook Pro equipped with an M4 Max chip or a more powerful processor. This means the AI processes data on the user's device rather than sending audio to a remote server. The team behind the project, led by Ragavan Thurairatnam, published sample recordings to demonstrate the system's capabilities. In one clip, the model engages in a dialogue that appears mostly natural, though it initially confuses the speaker with a person named Ronaldo. Other samples show the model drifting into tangents about film directors and expressing what the website describes as teary gratitude during a conversation about a trip to Paris.

The second point concerns the training methodology. Block describes Dessa as a true speech-to-speech model. The developers note that current voice assistants often flatten human speech into text, which they argue causes a loss of timing and context. By training only on audio, the model learns language in a way similar to how infants acquire speech. The team explains that babies learn to talk without seeing written words, and they wanted to replicate that process in artificial intelligence. They acknowledge that achieving this without text data is difficult but possible.

The third element is the safety argument. The team argues that many AI systems are deaf to the emotional tone of human voices. They present a scenario where a person is in serious pain, but the AI continues its routine because it cannot detect distress. Block positions Dessa as a step toward addressing this gap. The model is intended to be more responsive to the emotional cues embedded in speech, which the developers view as a critical aspect of human-machine interaction.

Block has not provided a timeline for a wider public release. The company plans to give early access to a small group of AI safety researchers. These researchers will evaluate the model's performance and safety before it reaches a broader audience. The announcement included several audio clips on the project's website, including one where Dessa is paired with OpenAI's GPT-Live-1 voice model. In that specific exchange, the audio sounds more natural than in other samples. The project is driven by what the team calls curiosity, a desire for better human-AI understanding, and a focus on safety.