Fish Audio Introduces Live 3: A Full-Duplex Voice AI for Real-Time Conversation

Fish Audio released Live 3 on November 10, 2026, a full-duplex voice AI model that continues listening while speaking, enabling more natural and responsive real-time conversations.

Live 3's ability to process input continuously during output enables it to adapt to conversational nuances such as pauses, acknowledgments, and mid-sentence corrections. This approach aims to create more fluid and human-like interactions with AI systems.

Understanding Full-Duplex Voice AI

Traditional voice AI models often operate in a half-duplex mode, meaning they either listen or speak, but not both simultaneously. This can lead to unnatural pauses and a less responsive conversational flow. Fish Audio's Live 3 addresses this by implementing a full-duplex system.

The core innovation of Live 3 lies in its capacity to listen while generating speech. This allows the AI to react to user input in as little as 186 milliseconds. For instance, it can interject brief confirmations or acknowledgments while formulating a more comprehensive response, mirroring human conversational patterns.

Key Capabilities and Training Methodology

Live 3 is engineered to handle common conversational challenges, such as mid-order corrections. An example provided illustrates its ability to process a change like "two pizzas... no, three" without requiring the user to restart their request or for the AI to reset its understanding.

According to CEO Rissa Cao, Fish Audio trained Live 3 on extensive real human conversations. This methodology allowed the model to learn natural conversational timing and dynamics organically, rather than relying on pre-programmed rules. The training corpus included a diverse range of dialogues, totaling 2,100 hours of stereo audio across 195,257 interactions. These dialogues covered everyday conversations, knowledge retrieval, tool use, and scenarios where the AI might abstain from responding.

Availability and Language Support

Fish Audio has confirmed that Live 3 supports the English language. The model is available through multiple channels, including the Fish Agent Platform, via API for integration into various applications, and for on-premises enterprise deployment. This broad availability aims to make the advanced conversational capabilities accessible to a wide range of developers and organizations.

Implications for Conversational AI

The introduction of Live 3 represents a step forward in the development of conversational AI. By enabling AI systems to listen and speak concurrently, Fish Audio aims to reduce the latency and artificiality often associated with AI interactions. This could lead to more engaging and efficient user experiences across various applications, from virtual assistants to customer service solutions.

The model's ability to learn conversational timing from real human interactions, rather than being explicitly programmed, suggests a more adaptive and nuanced approach to AI dialogue generation. This could pave the way for AI systems that are not only more responsive but also better at understanding and participating in complex human conversations.

Conclusion

Fish Audio's Live 3, released on November 10, 2026, introduces a full-duplex voice AI model designed to create more natural and responsive conversational experiences. By listening while speaking and reacting in milliseconds, Live 3 aims to overcome common limitations of previous voice AI systems. Its availability across various platforms positions it as a significant development for the future of AI-powered interactions.

Sources

• Context Spanning: A Communication Framework for Full-Duplex Speech Models and External LLM Backends
• Fish Audio Drama 3 Technical Report
• [2610.18010] Fish Audio Drama 3 Technical Report
• Fish Audio Drama 3 Technical Report
• fishaudio/drama-3 · Hugging Face