Shizuku AI puts its VTuber on an iPhone, no cloud needed

Shizuku is normally a YouTube character that reads viewer comments and answers in her own words. The prototype flips that setup: instead of typing a comment, a user speaks toward the phone, the speech is converted to text, an LLM decides the reply, and voice synthesis reads it back as Shizuku. To keep responses quick, the team leaned on the iPhone's GPU and its Neural Engine, the part of the chip that handles AI calculations.

The trade-off is explicit. Because the conversation runs on a lightweight language model that fits on a smartphone, the team says the prototype's back-and-forth is weaker than the regular Shizuku. The announcement frames the experiment as a check on how natural a phone-only call can feel at a time when LLMs are growing larger.

A separate but related release arrived on November 8: Shizuku-Duplex, a Japanese speech dialogue model for AI characters built on a full-duplex approach that lets it listen while speaking. It processes audio every 80 milliseconds, and searches or complex requests are handed off to another model in the background so the conversation can continue while a lookup is pending. In the company's own Japanese evaluation, it stopped speaking within 2 seconds of an interruption 89% of the time, above the 82% for the comparison GPT-Live-1.