How AI voice acting for games actually works (the bark, not the cutscene)

The search box sold you a cast

Type AI voice acting, game voice over, or character AI voice and the pages that come back treat every line as one job: pick a persona, hit generate, drop a folder of takes into Unity. That is a clean demo. It is a bad map if you needed to hear eight hundred guard barks in a playtest that is still rewriting itself.

Most of the work we see is not a cutscene. It is a bark sheet: short, regenerable lines so designers can test pacing, triggers, and whether the merchant sounds bored. Cutscenes and hero VO are a different job. They carry player trust, store disclosure, and often a human in the booth. Mixing those jobs is how a temp take becomes canon, then a Steam review.

The product page for this is AI voice acting for games. This post is the craft version: the three-job split, the cue that actually changes a bark, and the limits we print on purpose.

Three jobs, not one generate button

People type the same phrase for three products. Mixing them up is how a placeholder voice ships, or a lead performance gets treated like a spreadsheet.

Placeholder is the honest default

The useful sentence is the one on the Game Dev page: waiting on final VO blocks iteration. You need to hear the line in context, then change the line. Commissioning a session for dialogue that will move again is how budgets die in pre-alpha.

Generated temp audio is for that gap. It is not a claim that the machine replaced the actor. Teams that stay sane treat the file as scaffolding. QA will quote whatever is in the build. If you leave scratch in long enough, it stops being scratch.

When the script locks, you still get a choice: swap in a human, or keep the generated take. Commercial rights are the same either way. The choice is the point. We will not sell "replace your cast" as the product.

Barks are a numbers problem

A bark is short. House math on the Game Dev page puts a typical one around 60 characters. An 800-line sheet is not an audiobook. It is a re-render problem: combat tuning changes a trigger, and twelve guards need new lines by Friday.

That is why volume drafting still matters. Kokoro is the unlimited draft engine: a plain read, fast enough that rewriting the words is cheaper than arguing about the take. When a line has to sound alarmed instead of bored, move that line, not the whole sheet, onto the engine that acts direction.

On this roster that engine is Gemini Flash. It is the only one that follows bracketed cues. The Game Dev worked example is the same voice going [bored], [alarmed], and [angry] on one character, which is what a faction sheet needs before anyone books a booth.

Cues on the line, not a vibe slider

Most game-voice pages offer an emotion dropdown and call it acting. A dropdown is a style. A cue is a stage direction on the sentence that needs it.

Write the bark the way a script supervisor would: GUARD 2: [alarmed] Wait. That crate just moved. Keep the cue short. One mood per line. If you stack three feelings on a six-word bark, you are directing a cutscene in a cell that cannot hold it.

Cues are ignored by the rest of the roster. Grok Voice will read [angry] out loud as brackets. So will Kokoro, MAI Voice 2, and Zonos. We already wrote the longer version in cued TTS vs flat reads and the short definition in what is a TTS cue. The studio badges Acts [cues] or Plain read so this is not a surprise.

Personas are not clones

A character that has to stay the same across fifty lines needs a voice that does not drift. Grok Voice is the persona engine here: Eve, Ara, Rex, Sal, Leo. Pick one per character and leave it. That is a cast choice, not a cue.

It is also not voice cloning. Cloning a specific actor, or a temp you recorded in a closet, is still coming soon, English-first, with a consent step. We do not imitate a real person without that attestation, and we will not advertise a clone picker we have not connected. Until then, roster voices and personas are the honest path.

One voice per character keeps the build coherent. Swapping engines mid-faction because a demo sounded prettier is how the night-shift guards become two people.

What to send where

This is not a second copy of pick the engine for the job. It is the game-dev slice of that map, grounded in the live Game Dev recommended engine and the roster traits.

Files in the pipeline you already have

The live path is Text to Speech. Paste the line, pick a voice, generate, export MP3 or WAV. There is no wrapper, no SDK, no per-seat middleware. The file drops into Unity, Unreal, Godot, or a folder named temp_vo the same way any other take does.

We do not lip-sync. We do not drive visemes. We do not sit inside Wwise. If your pipeline needs those, you already have a tool for them. A studio WAV is the honest output of this product.

Beds under a trailer or a pause menu are a second export. Generate the voice here, generate a quiet instrumental in Sound & Music, mix them in the editor. Do not ask the music model for a footstep. We already said that in how AI sound effects actually work.

Where it still fails

A fluent bark can still be the wrong file. These are the misses we see when people treat the search phrase as a spec.

• A cutscene directed like a bark, or a bark directed like a cutscene.
• Cues on an engine that reads them aloud.
• Three moods stacked on six words.
• A different engine per line in the same character.
• Scratch audio left in until QA, and then players, treat it as final.
• A clone of a real actor, or a celebrity bit, with no consent path (we do not offer that, and acceptable use would not either).
• Expecting mouth shapes, real-time dialogue, or an in-engine API from an exported file.

If the line ships, say so

Placeholder audio you replace before release is a production trick. Lines that stay in the build are content players hear. Those are different filings.

As of September 2026, Valve's Steamworks Content Survey asks you to describe generative AI used to create content players consume. The survey splits pre-generated assets that ship with the game from live-generated output while the game is running. Sound is in the examples, alongside art and narrative. Read Valve's page before you file. We are not Valve, and store rules move. Verify before you rely.

Player trust is a separate problem from the form. Reviews will call a take AI even when it is not. If you ship generated VO, say so in the place your store already asks. If you do not, do not hide a temp folder in the pak. Neither posture is a marketing slogan. It is the receipt.

Honest limits

This is not a union-rules desk and not a middleware vendor. We generate files. You own the exports: commercial use, no watermark, same as the rest of the studio (you own what you make here).

No lip-sync. No visemes. No Wwise or FMOD plugin. No live NPC chatter API. The console API tab is still labeled soon; we will not pretend a public game endpoint exists.

Voice cloning is not a live path today. When it ships it will be English-first with a consent attestation. Until then, use roster voices and personas.

Premium generation sits inside a flat monthly allowance (published free-tier premium budget: 10,000 characters, about 10 minutes by house math). Kokoro drafting stays unlimited so you can rehearse a sheet without burning keeper budget on every micro-edit. Paid plans raise the flat allowance; they do not put you on a per-character anxiety meter.

Voice one sheet this week

The shortest proof is a faction you already have in a spreadsheet, not a roundup of vendors.

• Pick twenty barks for one character. Keep them short.
• Draft the words on the unlimited engine until the lines are actually the lines.
• Move the ones that need a mood onto Gemini Flash. One cue per line.
• Export WAV. Drop the files into the build with the names your pipeline already uses.
• Play the night shift. Change the three lines that do not land. Regenerate those three, not the sheet.
• If you are looking at a two-minute cutscene, stop. Book a person, or wait until the scene is locked and you still mean to ship a generated take on purpose.

Open the studio, not the Oscar

If you came here from an AI voice acting or game voice over search, the honest product is placeholder dialogue and bark sheets you can hear in the build this week. Open Game Dev, or start in Text to Speech if you already have the lines.

When you want the roster traits in one place, stay on engines. When you want the full studio, create a free account. Written to be heard includes the guard who has nothing but rats on third watch, as long as we do not lie about what that file is for.