Shizuku-Duplex: Japan's first full-duplex voice model that listens while it talks
Shizuku AI has unveiled Shizuku-Duplex, a voice model that holds a conversation in Japanese the way people do — listening while it talks, nodding along, and letting you interrupt. Apart from OpenAI's GPT-Live, it is the first full-duplex model capable of intelligent dialogue in Japanese, and it answers in about half a second.
- Announced: November 8, 2026, by Shizuku AI — a technical blog post by Ayuto Tsutsumi, Kohei Ohno, Yuu Jinnai, and Akio Kodaira.
- What it is: a full-duplex Japanese dialogue model — the first capable of intelligent conversation in Japanese besides OpenAI's GPT-Live.
- Tick: 80 milliseconds; on every tick the model listens to the user and produces its own speech, with TensorRT and Triton kernel optimisations keeping each tick inside its budget.
- Reply speed: 0.57 seconds median after the end of speech, against 0.53 seconds for GPT-Live, 1.27 seconds for Gemini Live, and 1.61 seconds for GPT-Realtime.
- Interruptions: stops within 2 seconds of being cut off in 89% of cases, ahead of GPT-Live at 82%.
- Tool calling: the first Japanese full-duplex model with tool calling besides GPT-Live; a background model runs web searches — 7.1 seconds in the Kyoto-trip demo — while the conversation continues.
- Accuracy: 79% on multi-turn question answering and 43% on web-lookup questions, against GPT-Live's 95% and 50%.
- Availability: no public release, API, or open weights announced; research and demos only.
Most voice AI still works like a walkie-talkie: you talk, it waits, it thinks, it answers. Shizuku AI, the Japanese startup behind the Shizuku companion AI, unveiled a model on November 8 that works more like a phone call. Shizuku-Duplex listens while it speaks — it can murmur a backchannel while you are still talking, let you cut in mid-sentence, and start replying before you have finished. Apart from OpenAI's GPT-Live, which arrived in July 2026, Shizuku AI says it is the first full-duplex model capable of intelligent dialogue in Japanese.
That is a narrower claim than it sounds, and the company is careful about it. Japanese full-duplex models already exist — J-Moshi and LLM-jp-Moshi-v1, both built on Kyutai's Moshi — but Shizuku AI's point is that they could hold a conversation, not a smart one. In its own evaluation, those two scored 3.4 and 2.9 out of 10 on chit-chat naturalness, against 6.2 for Shizuku-Duplex and 7.6 for GPT-Live. The gap is the story: the model is not just speaking Japanese, it is keeping up with a Japanese conversation.
An 80-millisecond heartbeat#
Shizuku-Duplex runs on the same clock as the person it is talking to, with every 80 milliseconds counting as one tick. On each tick it listens to the user's voice and produces its own speech, writing out what it is about to say as text slightly ahead of the audio — an "inner monologue" that lets it decide for itself when to start talking and when to stay quiet. Each tick has to finish processing within 80 milliseconds, which Shizuku AI achieved with inference optimisations using TensorRT and Triton kernels.
The architecture is an LLM-based Transformer with a streaming audio encoder, and it produces three outputs per tick: audio, text, and an action — the action head picks a Live2D motion so the companion character moves in a way that matches what it is saying. The result, in the company's demo calls, is a model that says "mm-hmm" while you talk and stops when you interrupt, rather than ploughing on through your sentence.
Two models, one conversation#
The harder problem is what happens when the conversation needs a fact. Looking something up, thinking through a complex problem, or calling a tool is difficult to do while holding a conversation at the same time — so Shizuku-Duplex runs on two models: an interaction model that handles the conversation, and a background model that works through difficult requests behind the scenes, asynchronously, with access to tools such as web search.
When a lookup is needed, the interaction model emits a special DELEGATE token. In the demo, a user asks to keep a Kyoto trip to about 100,000 yen for two people; the background model takes the request and searches the web for 7.1 seconds while the interaction model keeps chatting — "Matcha sweets sound great. Within 100,000 yen… let me look for a few plans" — then delivers the answer in its own words once the result arrives. Shizuku AI says this makes it the first Japanese full-duplex model capable of tool calling, except for GPT-Live. The design echoes Thinking Machines Lab's Interaction Models, which the blog cites as its basis.
Long conversations get a second mechanism: auto-compaction. Because a full-duplex model's context keeps growing even during silence, responses degrade as a chat gets longer. Shizuku-Duplex summarises the conversation roughly every five minutes and rebuilds its context, with the switch processed inside a single tick so the dialogue never pauses.
The numbers — and the caveats#
There is no public benchmark that can evaluate Japanese full-duplex models — the two existing ones, M3-DuplexBench and Full-Duplex-Bench-JA, have not published their Japanese evaluation data — so Shizuku AI built its own, derived from the English Full-Duplex-Bench and M3-DuplexBench, and scored the content with Claude Opus 5.5 as judge. That is the caveat to carry through every number that follows: the benchmark is the vendor's own, and the judge is a rival lab's model.
On the numbers the company reports, Shizuku-Duplex was best among the full-duplex models on five of the six turn-taking metrics. Its median reply came 0.57 seconds after the end of speech, close to GPT-Live's 0.53 seconds and well ahead of Gemini Live at 1.27 seconds and GPT-Realtime at 1.61 seconds — though the two non-full-duplex models were measured with network round-trip time included, and Shizuku-Duplex was not. It stopped within two seconds of an interruption in 89% of cases, ahead of GPT-Live at 82%, and stayed on topic after an interruption 82% of the time, against GPT-Live's 84%.
The quality gap is wider than the timing gap. On multi-turn question answering Shizuku-Duplex scored 79% against GPT-Live's 95%; on web-lookup questions, 43% against 50%. Chit-chat relevance came in at 6.5 out of 10 against GPT-Live's 8.5. In other words, the conversation mechanics — the listening, the timing, the interruptions — are close to the frontier, while the answers themselves still lag it. The model is better than GPT-Live at ignoring speech that is not meant for it: it responded to background speech in 75% of cases, against 92% for GPT-Live and 95% for J-Moshi-ext.
What it is for — and what is not shipping#
Shizuku-Duplex is a step toward Shizuku, the empathetic companion AI the company is building — an AI that lives alongside its user, responds to how they are doing, and openly expresses its own feelings. That is a different goal from GPT-Live, which Shizuku AI describes as a tool designed to solve the user's problems. The companion framing explains the Live2D motion head and the emphasis on presence: nodding, pauses, and a face that changes with the conversation are the point, not decoration.
What is not shipping is anything you can use. The announcement is a technical blog post with recorded demo calls and evaluation tables — no public release, no API, and no open weights have been announced. The stated next steps are adding vision input and expanding support to other languages. The broader context is a crowded month for live voice: ElevenLabs shipped two speech-to-speech models for live voice agents in October, and Mistral's Voxtral 3 arrived as a full voice-agent stack covering 50-plus languages — but both are built for agents and customer service, not for a companion that talks back in Japanese.
What to watch#
Three things to watch. First, whether Shizuku AI releases weights, an API, or even a public demo — until then, the benchmark numbers are the company's own and nobody outside it can reproduce them. Second, whether the two-model design holds up outside the demo: a 7.1-second web search is fine when the model can stall gracefully, but a hard question that needs several lookups will test how long the small talk can stretch. Third, the Japan angle: Meta picked Japan as the first market outside North America for its Muse agent in October, and Japanese is now one of the most contested languages in voice AI. A model that can actually converse in it — interruptions, backchannels, and all — is a foothold, if it ever ships.
Until then, the honest summary is short: a small Japanese startup built a voice model that talks at the same pace as a person, measured it against the best in the world, and came second on quality and first on manners. The question is whether anyone outside the company will ever get to talk to it.
Sources: Shizuku AI, ‘Shizuku-Duplex: Japanese Full-Duplex Model for Natural Conversation’ (November 8, 2026), here; HCI Insights, ‘HCI Insights for November 8, 2026’ (November 8, 2026), here. Cover and inline images are AI-generated illustrations.
Frequently asked questions
What makes Shizuku-Duplex different from a normal voice assistant?
A normal voice assistant waits for you to finish, then runs speech recognition, a language model, and speech synthesis in sequence. Shizuku-Duplex listens while it speaks: it works in 80-millisecond ticks, so it can nod along with a backchannel, let you cut in mid-sentence, and start replying without waiting for the end of your turn.
How does it search the web without pausing the conversation?
It runs two models at once. The interaction model keeps the conversation going while a background model handles hard requests through a tool harness with web search. When a lookup is needed, the interaction model emits a DELEGATE token; in the demo, a Kyoto trip search took 7.1 seconds while the model kept chatting, then delivered the answer in its own words.
Can I try Shizuku-Duplex, or download the model?
Not yet. Shizuku AI announced the model with blog demos and evaluation results, but has not announced a public release, an API, or open weights. The model is a step toward Shizuku, the company's empathetic companion AI; vision input and support for more languages are the stated next steps.