Shizuku AI introduces full-duplex voice model Shizuku-Duplex
Shizuku AI’s Shizuku-Duplex listens while speaking in 80 ms ticks. The company reports a 0.57-second median reply, excluding network time.
Shizuku AI, a Japanese company, announced Shizuku-Duplex on November 8, 2026: a Japanese-language full-duplex voice model designed to listen while it speaks. Its 80 ms figure is the model’s processing tick, not the wait from a user’s last word to a reply. The company separately reports a 0.57-second median for that measure, based on server-side recordings that exclude network round trips.
Key points
- Shizuku-Duplex processes audio in 80 ms ticks and aims to generate speech while it continues listening.
- The company says the model led full-duplex systems on five of six turn-taking measures in its Japanese evaluation; these are vendor-run results.
- The same evaluation reports an 89% rate of stopping within two seconds of an interruption, but only 17% and 25% rates for avoiding a takeover when the user is talking to someone else or background speech is present.
Background
Many voice assistants still run a relay: speech recognition turns a user’s words into text, a language model drafts a reply, then speech synthesis reads it aloud. That division makes components easy to swap, but it also compresses tone into a transcript and leaves a pause between turns. Full duplex moves the hard question from how quickly a transcript arrives to when a system should speak, stop, or stay quiet.
Our coverage of MAI-Transcribe-2-Streaming followed the streaming transcription side of that chain. Block’s Dessa is another full-duplex speech-to-speech model, with a different emphasis on its training route and local execution. Neither is a direct ranking comparison under Shizuku AI’s evaluation. Shizuku AI focuses on the rhythm of two-way interaction, interruption handling and Japanese conversation, while saying a background model can take on complex requests asynchronously.
The facts
Shizuku AI describes the core as an LLM-based Transformer. It handles audio features in 80 ms ticks and produces speech, text described as an “Inner Monologue,” and motion actions. The company says each tick must finish within the next 80 ms window and that it optimized the system with TensorRT and Triton kernels. This describes the processing cadence; it does not establish an 80 ms user-facing response time.
A separate background model handles harder requests and tools such as web search. The interaction model can keep talking instead of waiting for the search to finish, then use its result in a later response. For long sessions, the system summarizes conversation history about every five minutes; the company says it can switch to the refreshed context within one tick. The announcement does not specify public API access, downloadable weights, pricing or a waitlist, so developers cannot tell from this post alone how to try the model.
The company blog reports these selected evaluation figures:
| Measure | Reported result | What it means and its limits |
|---|---|---|
| Processing tick | 80 ms | Audio processing and generation window, not end-to-end reply latency |
| Turn response | 0.57 s median | From the user’s turn ending to the model’s reply; server recordings exclude network round trips |
| Stop after interruption | 89% | Share of interruptions after which the model stopped speaking within two seconds |
| Avoiding unwanted takeover | 17% / 25% | No takeover when the user talks to someone else: 17%; no takeover with background speech: 25% |
The comparison includes Shizuku-Duplex, GPT-Live, J-Moshi-ext, LLM-jp-Moshi-v1, GPT-Realtime and Gemini Live. Shizuku AI says its model ranked first among full-duplex models on five of six turn-taking measures, while its conversational quality ranked behind GPT-Live. The team assembled the Japanese evaluation, transcribed speech with Haas-ja-v1.0 and used Claude Opus 5.5 to score content. The blog says the Japanese evaluation data for M3-DuplexBench and Full-Duplex-Bench-JA is not public; it does not say whether Shizuku’s own evaluation data is public. Latency comparisons also use different conditions: Shizuku-Duplex and GPT-Live were measured without network round trips, while the API measurements for GPT-Realtime and Gemini Live include network effects.
What others say
PC Watch reported the 80 ms tick, TensorRT and Triton optimization, and the 0.57-second median on November 9, framing the system as an architecture for backchannels, interruptions and simultaneous listening and speech. GIGAZINE covered how the system gives a short response while web search runs, then adds the search results, as well as periodic conversation summarization. Neither article describes a reproduction or an independent test. IT之家 also recapped the launch and the companion character 雫, but did not identify who ran the six-model evaluation or explain its method.
So the account currently rests mainly on the company blog and two Japanese outlets. Those reports help check the launch date and system description; they do not replace an external rerun of the dataset, scoring process or network conditions. The official blog says the Japanese evaluation data for M3-DuplexBench and Full-Duplex-Bench-JA is not public; it does not state whether the team’s Shizuku-specific evaluation data is public. We found no third-party reproduction in the material reviewed. The ranking, latency and interruption rates should be read as vendor-reported results, not independent findings.
Our take
For an AI companion, voice character or live support agent, “listening while speaking” only helps if the system also knows when to stop. The 89% two-second stop rate suggests it can respond to some explicit interruptions. But its 17% no-takeover rate when a user is speaking to someone else, and 25% with background speech, point to a harder product problem: people talk around these systems, not only into them. Builders should test conversational courtesy, multiple speakers and television or room noise separately from a clean demo.
The design is worth watching: a low-latency interaction loop keeps a conversation moving while an asynchronous path retrieves answers to harder questions. That could suit prototypes for voice products built around sustained interaction. But the announcement does not confirm an accessible API or public weights, and the Japanese evaluation data is not available for review. What can be checked today is the architecture and the company’s own results. Judging the experience will require public evaluation materials, latency measurements under the same network conditions, and an outside rerun in noisy and multi-speaker settings.