For years, talking to an AI has felt like a high-stakes game of 'wait your turn.' You speak, you pause, the AI processes, and then it responds. It’s functional, but it isn’t a conversation. Enter Dessa, the new full-duplex speech model from Block that promises to shatter that awkward silence and bring us closer to truly human-like interaction.
The Magic of 'Speech-Only' Training
What makes Dessa different isn't just how it sounds, but how it was born. Most voice AIs are essentially text models with a voice skin stretched over them. Dessa, however, was trained from scratch exclusively on recordings of human conversation. No text pretraining, no hidden LLM scripts—just raw audio.
By learning directly from how humans actually talk, Dessa has captured the 'invisible' parts of communication: the nuanced pauses, the emotional shifts, and the natural cadence that makes a voice feel authentic rather than synthesized.
True Full-Duplex Conversation
In tech terms, 'full-duplex' means the ability to send and receive data simultaneously. In human terms, it means you can interrupt the AI, and the AI can react to you in real-time without breaking character.
Because it doesn't have to translate speech to text and back again, the latency is remarkably low. In fact, reports indicate it can run faster than real-time locally on an M4 MacBook Pro. This removes the 'robotic lag,' allowing for a flow where interruptions and overlapping speech feel natural rather than glitchy.
The Future of Intelligence
While currently a research preview, Dessa signals a shift toward a more intuitive era of AI. We are moving away from 'prompting' and toward 'chatting.' If Block can scale this authenticity, the interface of the future won't be a screen or a text box—it will simply be a voice that listens and speaks exactly like we do.
Sources
Media
