The Japanese AI laboratory Shizuku AI has introduced Shizuku-Duplex, a full-duplex model designed for Japanese-language conversational AI. It is the second model of its kind to support real-time interaction, following OpenAI's GPT-Live. The system initiates its thought process without requiring a complete query from the user.
In performance evaluations, Shizuku-Duplex achieved the highest naturalness score among six tested models for Japanese language processing. Its response quality is ranked immediately behind GPT-Live. The architecture includes a foreground model for interaction and a background model for handling complex requests asynchronously.
The system utilizes a context compression mechanism that triggers every 5 minutes. It operates with an interaction frame rate of 80ms and supports three distinct output formats: audio, text, and physical actions. Shizuku AI is currently leveraging this technology to build a companion AI capable of continuous listening and emotional expression.
