Japanese AI Firm Shizuku AI Launches Full-Duplex Voice Model 'Shizuku-Duplex' for Real-Time Interaction
Shizuku AI released a full-duplex Japanese audio model designed for AI characters, achieving 0.57-second median response time via 80ms processing ticks.
Japanese AI startup Shizuku AI announced "Shizuku-Duplex," a full-duplex Japanese audio dialogue model, on November 8. The model is designed to enable natural human-AI interaction for AI characters, allowing them to listen and respond simultaneously, even starting replies before the user finishes speaking.
Unlike traditional cascaded systems that sequentially process speech recognition, text generation, and synthesis, Shizuku-Duplex uses a full-duplex architecture operating in 80-millisecond ticks. In each tick, it listens to user audio while generating its own. The system also features an asynchronous background model for tasks like web search, maintaining conversation flow while waiting for results.
According to Shizuku AI's internal evaluation, the model achieved the highest conversational naturalness among six compared models, with a median response time of 0.57 seconds after the user stops speaking. It stopped within two seconds of user interruption in 89% of cases. The company plans to expand language support and add image recognition capabilities.