Block has unveiled Dessa, a full-duplex speech-to-speech model trained solely on human conversation audio, though access remains limited and its capabilities raise unanswered questions.
Block has introduced Dessa, a voice artificial intelligence model that the company says was trained entirely on human conversation audio rather than on text or a pretrained language system. In its demonstrations, the model is shown handling spoken exchanges in real time, with laughter, pauses and changes in volume that are designed to mirror the rhythms of ordinary conversation. According to Block, the system works as a speech-to-speech model and preserves cues that are often stripped away when voice products first convert speech into text.
The company describes Dessa as a full-duplex system, meaning it can listen and respond at the same time rather than waiting for a speaker to finish before replying. Block says that approach allows it to overlap with a human speaker in a way that feels closer to natural dialogue. The model is also said to decide what to say, when to speak and how to sound, including aspects of voice and personality, and the company claims it can run locally on a MacBook Pro with an M4 Max chip or better.
The public demonstrations are not entirely polished, and that is part of the point. In one clip, the model initially mistakes its conversation partner for someone named Ronaldo before the exchange settles into a more ordinary flow. In another, a discussion about Christopher Nolan shifts unexpectedly into Bryan Singer’s Superman films. A separate example shows a user mentioning a surprise trip to Paris, prompting a response that the team described as sounding tearfully grateful. The website also includes a conversation between Dessa and OpenAI’s GPT-Live-1 voice model.
Block says the project is meant to explore how speech systems can learn from listening alone, in a way the company likens to how babies acquire language before learning to read. It also argues that a model which can detect distress, tone and other emotional signals in speech could matter for safety as voice agents become more capable and autonomous. For now, access is limited to a small group of AI safety researchers, and Block has not announced a public release date or API. The company has also raised open questions about how a speech-trained model acquires facts, learns languages and incorporates information introduced during conversation.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
