hikomi Open the app

AI audio

Every voice in hikomi is AI-generated. On the website, and in the desktop app from version 0.5.4, the audio of each spoken reply, voice preview and Demo sample, and of the short lines played while a voice wakes up, carries a hidden, machine-readable mark saying that it is AI-generated audio made with hikomi. So does the sound of the videos about hikomi in which a character speaks. The mark is the same for everyone and holds nothing about you.

Audio check

Check a clip

WAV, MP3, Ogg/Opus, FLAC, M4A or WebM; mono or stereo, up to 8 MB. At least 1 second of audio is needed. Only the first 20 seconds are checked. The file is sent to our server only to look for the mark, and it isn't kept.

Choose a clip to begin.

What a check can tell you

A mark can be weakened or lost by editing, speeding up, re-recording through a speaker, or heavy compression. "No mark found" does not prove a clip was made by a person. A clip from a desktop app older than version 0.5.4, or from a copy of a video about hikomi made before 15 October 2026, may not carry the mark.

For researchers

The watermark uses Meta's AudioSeal (MIT licence), with the 16-bit message 0x484B (HK). A clip counts as marked when the detector's score is at least 0.9 and it reads that message; below 0.1 counts as no mark found, and anything else as unclear. In stereo both channels must agree, and the lower score is shown.

Marked Ogg and MP3 files also carry the tag AI_GENERATED=hikomi;mark=audioseal. The same check answers at POST /api/voice/check, with the file as the raw request body and the same limits.