Voice AI just got more human: new model listens and talks at the same time

Fish Audio has launched Live 3, a full-duplex voice model designed for real-time conversational agents. Unlike traditional systems that require hand-coded rules for turn-taking, Live 3 continuously listens while speaking, allowing it to handle natural interruptions, backchanneling (like "mm-hmm"), and mid-sentence corrections without breaking the conversation flow. The model was trained on real human dialogues to learn conversational timing directly from data.

For developers building voice assistants and interactive agents, this represents a shift from rigid, rule-based pipelines to more fluid, model-driven interactions. Live 3 can decide to respond in as little as 186 milliseconds for short acknowledgments, enabling more natural and responsive conversations. The takeaway: the next generation of voice AI is moving beyond simple speed to master the nuanced rhythm of human dialogue.