Dessa: Conversation Model trained only on speech

Dessa is Block’s speech-to-speech AI, designed to converse in real time without converting speech into text. Its developers say it was trained from scratch on human conversation alone, with no text data or pretrained models, so it learns from vocal sound and timing. The model listens and speaks simultaneously, can interrupt or offer brief acknowledgments, and responds to features such as emotion, sarcasm, laughter, pauses, and changes in emphasis. It runs faster than real time on a MacBook Pro M4 Max or better, and the team says it was trained using a single NVIDIA GB200 NVL72. Demonstrations show it laughing mid-sentence, adjusting its volume, recalling information, responding empathetically, and speaking with varied intonation; the recordings are presented as live conversations rather than scripted performances.

The project explores whether machines can acquire language and conversational knowledge from sound, and how speech-only learning compares with text-based systems. Its creators argue that text transcription discards emotional and social cues, making voice assistants feel unnatural and potentially unable to recognize distress. They see Dessa as a step toward more natural interaction and better safety, while also opening research questions about learning language, facts, and new languages from audio. Because the model represents new territory, Block is inviting a small group of AI safety researchers to examine possible risks. The project comes from a small research team at Block with experience in deep learning and speech generation; access is currently offered through a waitlist and a research-access application.