translated from Japanese. Please refer to the original ASCII.jp article for the authoritative version
On <a href="https://comicvibe.com/rj-mitte-honored-by-co-stars-at-celebration-of-life-on-november-7/” title=”RJ Mitte Honored by Co-Stars at Celebration of Life on November 7″>November 5th, the Shizuku AI Technology Department released a prototype demo of voice conversation with AI VTuber “Shizuku” using only an iPhone. Voice recognition, AI response generation, and speech synthesis are all processed within the iPhone without relying on the cloud.
Shizuku is an AI character that reads viewer comments during YouTube streams and responds in her own words. In this prototype, instead of sending text comments, users speak directly to the smartphone. Voice recognition transcribes what is spoken, a Large Language Model (LLM) formulates a response, and speech synthesis reads it aloud in Shizuku’s voice.
According to the explanation, the iPhone’s GPU and Neural Engine, which handles AI computations, are utilized to speed up responses. The conversation uses a lightweight language model designed to run on smartphones, meaning its conversational ability is somewhat reduced compared to the usual Shizuku.
Additionally, on November 8th, Shizuku AI also announced “Shizuku-Duplex,” a Japanese voice dialogue model for AI characters. It adopts a method called “full-duplex,” which allows users to hear the other party’s voice while speaking.
The dialogue model processes audio every 80 milliseconds. Since searches and complex requests are handled by a separate model in the background, conversations can continue even while waiting for information.
In the company’s Japanese evaluation, the percentage of times the AI stopped speaking within 2 seconds of being interrupted was 89%, surpassing the 82% of the comparative GPT-Live-1.
We experimented with chatting with Shizuku using just an iPhone! 📱 As LLMs become larger, our staff prototyped this out of curiosity, wondering “how naturally can we converse using just a smartphone?” Because of that, the “brain” is a lightweight size designed for smartphones, so Shizuku is a bit less sharp (or perhaps ‘clumsier’?) than usual 😂 https://t.co/UX1kj7Jg7Y — Shizuku AI技術部 (@ShizukuResearch) November 5, 2026
