Japanese AI laboratory Shizuku unveiled the Shizuku-Duplex full duplex model on November 9, claiming it is the second such model after GPT-Live
The model ranks highest for Japanese conversational naturalness among six tested models with response quality second only to OpenAI GPT-Live, using a foreground interaction model plus asynchronous background model with context compression every five minutes and an 80 millisecond interaction frame
On November 9, Japanese AI laboratory Shizuku unveiled its Shizuku-Duplex model, announced locally the previous day, claiming it is the second full duplex model supporting intelligent Japanese conversation after GPT-Live, able to start its thinking flow without waiting for a complete question. In Japanese, Shizuku-Duplex has the highest conversational naturalness among the six tested models, with response quality second only to OpenAI GPT-Live.
Shizuku-Duplex splits into a foreground interaction model that handles dialogue and a background model that processes complex requests asynchronously, assisted by context compression every five minutes. The system treats 80 milliseconds as one interaction frame and supports three output modes: audio, text monologue and motion.
Building on the model, Shizuku is developing a companion AI that always listens, responds to how the user feels and openly expresses its own feelings. The key difference of full duplex voice models is that they begin generating a reply without waiting for the user to finish an entire sentence, producing a rhythm closer to human conversation, which is a direction few products have covered in the Japanese market so far.