Shizuku AI Develops Real-Time Japanese Voice Model Shizuku-Duplex
Shizuku AI announced Shizuku-Duplex, a Japanese full-duplex voice model that runs real-time inference every 80 milliseconds. The model can interject backchannels while a user speaks and respond before the user finishes.
1 source 25m ago
Reporting from 1 source: GIGAZINE.
Why it matters
Shizuku AI is staking its AI VTuber business on conversation pacing rather than raw model scale, and its own benchmark places Shizuku-Duplex behind GPT-Live-1 while well ahead of existing Japanese full-duplex models.
Key Facts
- Shizuku AI announced Shizuku-Duplex, a full-duplex model for real-time Japanese voice conversation.
- Shizuku-Duplex runs real-time inference every 80 milliseconds.
- The model can interject backchannels while the user is speaking and respond before the user finishes.
- Shizuku AI positions Shizuku-Duplex as the first full-duplex model besides GPT-Live capable of intelligent Japanese dialogue.
- Shizuku-Duplex summarizes conversation history every 5 minutes for stable output in long conversations.
It is positioned as the first full-duplex model besides GPT-Live capable of intelligent Japanese dialogue, and Shizuku AI reports performance second only to GPT-Live-1 among full-duplex models.
Shizuku AI built the model for its AI VTuber Shizuku, and a posted conversation example has the persona interjecting short backchannels, answering without a pause after the user stops, and saying "please wait a moment" before a detailed reply. Before audio comes out, Shizuku-Duplex writes the dialogue line as text in an "Inner Monologue" step. Inference optimization with TensorRT and Triton kernels supports the 80-millisecond cycle.
For questions that need a web search, the model gives a short spoken answer while the search runs, then delivers the detailed answer without breaking the flow. Conversation history is summarized every 5 minutes to keep long sessions stable. Shizuku AI says it will continue work on vision and motion.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google