All news

Shizuku AI unveils Japanese voice model that listens and speaks simultaneously

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

Voice SI is moving closer to natural conversation: it can keep talking while searching for information instead of falling silent while processing a request.

Shizuku-Duplex processes dialogue in 80 ms steps: at each step, the model receives the other person's speech and generates its own text and audio tokens. This allows it to produce backchannels, choose pauses and accept interruptions without waiting for the speaker to finish. The system is split into a live interaction model and a background model. The latter calls tools asynchronously, such as web search, while the former continues the conversation. The developers also added context summarization approximately every five minutes for long conversations. In Shizuku AI's own test, the model produced the most natural dialogue among six systems evaluated and ranked behind only GPT-Live for response quality among full-duplex models. The median pause before a response was 0.57 seconds, compared with 0.53 seconds for GPT-Live; measurements for these two models excluded network latency.

Sources 1