Voice and NPCs
Shizuku-Duplex: a full-duplex voice model for Japanese
Shizuku AI Labs has built a voice model that listens and speaks in 80 ms steps, backs off when interrupted and answers in about 0.57 seconds. Apart from OpenAI's GPT-Live, it is the first full-duplex model that can hold an intelligent conversation in Japanese, and a useful template for talking NPCs.
On 8 November, Shizuku AI Labs introduced Shizuku-Duplex, a Japanese dialogue model that listens and talks at the same time. The lab says that, apart from OpenAI’s GPT-Live, it is the first full-duplex model capable of intelligent conversation in Japanese.
Most voice AI waits for you to stop, then runs speech recognition, a language model and speech synthesis in turn. A full-duplex model does none of that waiting. It nods along with “mm-hm”, answers before you finish and stops when you cut in.
How it works
Shizuku-Duplex runs on an 80 ms clock. On every tick it hears the user and produces its own audio. It also writes the text it is about to say slightly ahead of the speech, which the lab calls an inner monologue. Whether silent or speaking, it keeps hearing the other person, so it decides for itself when to talk.
Three further design choices stand out:
- Two models. An interaction model holds the conversation. When a request is hard, it emits a
DELEGATEtoken and a background model works on it, with web search and other tools, while the conversation carries on. - Actions. Besides audio and text, a third output selects a motion that fits the character’s feeling and intent.
- Compaction. Roughly every five minutes the model summarises the conversation and rebuilds its context within a single tick, so long calls do not degrade or pause.
To meet the 80 ms budget, the team used TensorRT and Triton kernel optimisations.
The numbers
Shizuku AI built its own Japanese benchmark from the English Full-Duplex-Bench and from M3-DuplexBench tasks, and checked it by reproducing OpenAI’s published GPT-Live scores. Content was scored with Claude Opus 5.5 as the judge. These are the lab’s own results.
- Median time from end of speech to reply: 0.57 seconds, against 0.53 for GPT-Live. The lab measured both on server recordings, so network time is excluded.
- Best of the full-duplex models on most turn-taking measures in the lab’s chart, though GPT-Live starts talking too early less often. Shizuku-Duplex stopped within two seconds when interrupted 89% of the time (GPT-Live 82%).
- Second to GPT-Live on content. Multi-turn question answering reached 79% against 95%, and chit-chat naturalness scored 6.2 out of 10 against 7.6.
- It clearly beats the existing open Japanese full-duplex models, J-Moshi-ext and LLM-jp-Moshi-v1.
Access and limits
The post does not mention an API, a price or open weights, so treat this as a research announcement for now. The model is Japanese only, and has no vision input yet. The lab says it plans to add vision and more languages.
Shizuku AI is building the model into an AI companion character called Shizuku, and the demos use that persona.
Our take
For people building interactive worlds, the architecture matters more than the leaderboard. A small front model that stays in the moment, a slower model that thinks behind it and a motion output that lets the character show what it means is close to the blueprint for a talking NPC. Compaction tackles a problem every long-running character has: remembering the last hour without going quiet.
It also shows that full-duplex voice is moving beyond English. We cover the design problem in the end of the dialogue tree, and a similar split between a front and a background model appears in our post on MOSS-Live.