Block has announced a new artificial intelligence model named Dessa, which was trained from scratch using only human conversation. The company describes Dessa as a true speech-to-speech model where every piece of knowledge originated from human speech.
Users can run the model locally on hardware such as a MacBook Pro M4 Max or superior devices. Developers created the technology to satisfy curiosity, foster better human-AI understanding, and address safety concerns.
The team argues that babies learn to speak without seeing written words, and they believe current voice AI wastes the natural interface of speech. Because AI continues to gain independence while remaining largely deaf to vocal emotions, Dessa represents a step toward resolving safety issues.
A small group of AI safety researchers will receive early access to the model. Ragavan Thurairatnam shared a recording of a conversation between the model and a speaker, during which the AI initially confused the person for someone named “Ronaldo”.
During the interaction, Dessa participated in a guessing game over voice and held a conversation with OpenAI’s GPT-Live-1 voice model.
