Why AI voices suit visual novels
Visual novels are long. A modest one can have tens of thousands of lines, spread across routes and endings, and full voicing is often the most expensive part of production. That's why many are only partly voiced, or not at all. AI voices change the maths:
- Full voicing becomes possible for small teams, with performers' samples cloned with their consent.
- Partial voicing gets easier: key scenes, main characters or the opening chapter.
- Demos and trailers can be voiced before the script is final.
- Rewrites don't strand old audio: changed lines are re-made in minutes.
Setting it up
- One project per game; the cast is shared by every script in it.
- A script per route, chapter or scene file, whichever matches how your engine's script is split. Smaller scripts make review easier.
- Cast every speaking character, named exactly as your script names them, so
Aiko: You came back.goes to Aiko. - Export dialogue from your engine's script as
Name: linetext, one line each. Narration lines can go to a narrator voice, or be left unvoiced. - Generate in bulk with Generate all new lines, review, and save.
Connecting files to lines in your engine
Each line's take comes out as a numbered WAV file, with Lines.csv listing the number, speaker, text and
file. Most engines identify voice files by a line ID or a file name in the script:
- Ren'Py can play a voice file for each line with its
voicestatement, or find files automatically from each line's identifier with its auto-voice setting. It plays WAV, and Ogg or Opus are smaller for distribution. - Unity-based VN frameworks and other engines usually reference an audio clip per line; see importing into a game engine.
- Keep your own mapping from line number to your engine's line ID in a spreadsheet, and rename files to the IDs on import, so script reordering never breaks the links.
Casting and consistency
- Record each character's sample in their typical mood. A cheerful heroine's sample should sound cheerful.
- For a character's big emotional scenes, record an extra sample (
Aiko crying) and give those lines to it. - Keep each character's sample unchanged for the whole game. Saved lines keep their sound even if you change it, but new lines wouldn't match.
- Japanese, Korean and Chinese releases can be voiced in those languages too. See Japanese, Korean and Chinese.
Before release
- Get written consent for every performer whose voice is cloned, covering a commercial release. See the consent template.
- Disclose AI voices on your store page; Steam, for example, asks developers about AI-generated content.
- Credit the performers whose voices were cloned.