AI voices for VTuber content and character streams

VTuber content splits cleanly into two halves, and only one of them is a voice problem you can solve ahead of time. The live half -- the stream itself -- is you, on a mic, in real time. The pre-recorded half is everything else: the lore video, the intro sequence, the skit, the character short, the "meet the cast" clip, the birthday special where three characters who share one physical throat somehow have a conversation. That second half is scripted, and scripted multi-character audio is exactly what AudioProducer is built to render.

This guide is about the pre-recorded half. It is specific about what works, and equally specific about what this tool does not do, because the wrong expectation here wastes a weekend.

What this does, and what it is not

AudioProducer is not a real-time voice changer. It does not sit between your microphone and OBS, it does not pitch-shift you live, and it will not give your model a different voice while you are on stream. That is a different product category and you should buy it from someone who builds it.

What AudioProducer does is take a written script and render it as finished multi-voice audio, with a separate voice per character, music beds, ambient soundscapes and one-shot sound effects, in one pass. You paste the script, the AI tags who says what, you fix what it got wrong, and you download an audio file. For a lore video or a skit, that is the entire production pipeline between "script is written" and "audio is in my editor."

The practical consequence: if your character's voice has to react to chat, this is the wrong tool. If your character's voice has to say a line you already wrote down, it is the right one.

Start with the formats that are already scripted

Most channels have three or four pieces of content that are written before they are recorded, and those are the ones to move first:

• Lore videos. The backstory drop, the world explainer, the "what actually happened in the last arc" recap. Usually narration plus a few character lines in flashback.
• Intro and outro stingers. Fifteen to forty seconds, rerecorded whenever the branding changes, and painful to redo by hand for that exact reason.
• Skits and shorts. Two or three characters, a punchline, sixty seconds. These are where a solo creator hits the wall fastest, because one throat cannot hold three distinct voices convincingly for very long.
• Character introductions. New member of the cast, new OC, a guest who exists only as art. One voice, thirty seconds, needs to sound consistent forever after.

Pick the skit first if you have one. It is the format where the difference between one voice and four is most audible, and it is short enough to finish in a sitting.

Set the cast up once, then stop thinking about it

Paste the script into a chapter and run Auto-Assign Characters. It reads the scene and tags each line by speaker -- narrator, named characters, in-world labels. On a clean script with speaker labels already in it, it gets most of the way there on the first pass.

Then do the part it cannot do for you: decide which library voice belongs to which character, listen to them back to back on the Voices page, and lock the choice down. The voices are swappable at any time, which is exactly the trap. A voice you change in month three retroactively makes months one and two sound like a different channel. Write the mapping down somewhere outside the app -- character name, voice name -- and treat it as canon.

If you have a large cast, group characters into folders from the character editing menu. A four-person skit does not need it; a channel with two years of OCs does. There is more on the casting decision itself in giving each character a different voice and, once you have shipped a few episodes, keeping a character voice consistent across a series.

The narrator is a character too

This is the single most common mistake in lore videos. Auto-Assign hands every unlabeled line to the narrator, which is correct behavior and usually not what you want, because in a lore video the narrator is frequently in character -- it is the persona telling you about her own world, not a neutral documentary voice.

Decide up front which one you are making. If the narrator is the persona, the narration voice should be the persona's voice, and the emotional register of the writing should match. If the narrator is a neutral frame, give it a voice that is deliberately unlike every character voice in the cast, so the ear can tell frame from scene without being told.

Either way, re-tag the narration pass by hand after Auto-Assign. It is a minute of work and it is the difference between a lore video and a wiki article read aloud.

Write the timing into the script instead of editing it in later

Comic timing in a skit is pause timing, and pauses are cheaper to set before the render than to cut in afterward. There are four levels of control and you will mostly use two:

• Default paragraph pause -- set once in Edit Project, applied between every paragraph break. For a skit, keep it short. Half a second reads as conversation; a second and a half reads as a dramatic reading.
• Custom paragraph pause -- a per-paragraph override for the beat before the punchline. This is the one that does the work.
• Inline pause -- a silent gap dropped anywhere inside a line, for the character who trails off mid-sentence.

Set the beat overrides while you can still see the script, not after you are staring at a waveform. Pauses and dramatic timing in AI narration goes deeper on the numbers.

Sound design without opening a second app

Run Auto-Assign Sounds after the characters are tagged. It reads the scene and places music beds, ambient soundscapes and one-shot effects to match what is happening -- a storm scene gets thunder and wind, a transition gets atmosphere. As with the character pass, treat it as a first draft: it is generous, and a sixty-second skit usually wants less than it suggests.

For a lore video the ambience is doing real work, because there is no visual scene change to carry you between locations. A bed under the flashback and silence under the frame narration will do more for comprehension than any line of dialogue you could add.

Ship it as separate files

Render with one button and download per chapter, so each piece comes out as its own audio file. Structure your project accordingly: one chapter per skit, or one chapter per section of a long lore video. That gives your editor discrete assets to cut against the animation rather than one long track you have to slice by hand.

If the same content is heading to a video platform as narration over art, the workflow overlaps almost entirely with narrating YouTube videos with AI voices and running a faceless channel, both of which cover the upload side this guide skips.

The honest limitation, up front

Three things to know before you plan a schedule around this. It reads text you paste in or an EPUB you upload -- not a video file, not a subtitle track, not a screenshot of a Discord message. Nothing here happens live. And a rendered performance is a performance of what you wrote, so a flat script renders flat; the tool fixes the number of voices you have access to, not the writing.

The free plan is 1,200 words per month, which is roughly one lore video script or two short skits -- enough to find out whether the cast you imagined actually sounds like the cast you imagined, which is the only question worth answering in week one.

Cast one skit and finish it

Do not start by building the whole cast. Take a sixty-second script you already wrote, give it three voices, set two pause overrides and render it end to end. You will learn more about which voices belong to which character from one finished skit than from an afternoon of auditioning the library, and the mapping you write down that day is the one your channel keeps. The rest of the catalog, including every guide linked above, is in the full guide index.

When you have the script open, start a free project and run Auto-Assign on it. Then go straight to the narration lines, because that is where a character piece either holds together or falls apart.