Block introduces Dessa, a full-duplex speech model built solely on audio data

On October 23, 2026, Block’s AI research lab unveiled Dessa, a full-duplex speech model trained exclusively on human conversation. Unlike most voice assistants that rely on text-based systems, Dessa processes audio directly, converting speech to speech without using text or pretrained models. The model listens and speaks simultaneously, learning from real conversations to determine when to interject and when to respond with cues like 'yeah' or 'got it'. Dessa’s design avoids the traditional three-step process of speech recognition, text modeling, and speech synthesis, which Block claims loses emotional nuances like laughter and sarcasm. The lab has not disclosed its architecture, but it highlights the challenges of working with speech, which requires significantly more data than text. Block’s demo clips show Dessa laughing and adjusting volume, interacting with OpenAI’s GPT-Live-1 without human intervention. The project page invites comparisons with other models like Moshi, which uses a text-based approach with 7 million hours of pretraining audio and 1,016 H100 GPUs. Block encourages readers to explore its project page, compare it with related models, and engage on social media platforms.