CONNECTION ESTABLISHED FILE: #NCR_LIPS00001

▶ FILE #NCR_LIPS00001 // DECRYPTED · TOOL // OPEN SOURCE

LIPSYNC ANIMATOR

Hand it a recording and the words spoken in it, and it builds the mouth animation for that line. Free, open source, and there's a demo in the browser if you'd rather not install anything.

// CYBERPUNK 2077 MODDING TOOL · V0.1.0 · MIT · WINDOWS, OR JUST THE WEBSITE · NO GENERATIVE AI

THE PROBLEM // WHY THIS EXISTS

BORROWED MOUTHS SHAPE THE WRONG WORDS

A voiced line in Cyberpunk 2077 plays two things at once: the audio, and an animation that moves the speaker's mouth. CD PROJEKT RED made one for every line in the game. A mod that records new lines has none.

So modders borrow. You pick a vanilla line of roughly the same length and point your line at its animation. The mouth opens and closes while the character talks, which is better than a face frozen mid-sentence, but it is shaping somebody else's sentence, and the pauses land where that sentence's pauses were. Watch one closely and you can see it.

That was where things stood when these gigs started recording people. I went looking in October 2026 for anything that builds an animation from new audio and came up empty. WolvenKit has a feature in development, but it copies the animation a vanilla line already has, and its own wiki page says it does not create facial animation from a custom audio file. The voice-line databases list lines and subtitles and carry no lipsync data at all.

So this one builds the animation instead of copying one. It works out when each sound is spoken in your recording and puts the mouth in the shape the game's own animators use for that sound. Out comes an .anims file, the format the game already uses, which WolvenKit and the game read like any other.

METHOD // FIVE STEPS

HOW A MOUTH GETS BUILT

STEP 01

Find the sounds

A forced aligner, Montreal Forced Aligner, takes the recording and the words and marks where each sound starts and ends. English has 39 of them. It gets handed the words, so it is not guessing at them.

STEP 02

Look up the shape

Each sound has a mouth shape, averaged from what CDPR's animators actually did on that sound across 1,148 of the game's own lines. A sound looks different depending on its neighbours, so each has a version per kind of neighbour: 1,088 combinations. The dictionary is numbers only, no game files.

STEP 03

Put it 70 ms early

The mouth moves ahead of the voice: lips close for an "m" before you hear it. Measured across 204 game lines from six characters, the mouth leads the middle of each sound by 60 to 70 ms, the same for every speaker. Each shape goes in 70 ms early, then blends to the next.

STEP 04

Open the jaw with the volume

Louder speech opens the jaw a little more, 0.0018 per decibel. The jaw tracks loudness more closely than it tracks any single sound, which is also why the jaw is close even without the aligner.

STEP 05

Leave the rest of the face alone

Eyes blink about 1.2 s in, then every 1.3 to 2.1 s, as in the game's lines. Every other control and all the head bones stay at the character's own rest, so they keep their expression and their neck. The animation fades in 0.3 s before the first sound and out over the last 0.3 s, with the face at rest on the first and last frame. Without that last part the game opens the mouth once more after the final word.

An animation adds to the face rather than replacing it: every value is a change from the face already there, and 0 changes nothing. That is why one of these plays correctly on any head.

EVIDENCE // SCORED AGAINST CDPR'S OWN

HOW GOOD IT IS

Ten game characters were taken out of the dictionary, the tool rebuilt their lines from the audio, and each result was checked against the animation CDPR shipped for that same line. The number is a correlation: 1 means identical movement, 0 means unrelated. The last column is what mods do today.

Part of the mouthWith the alignerWithout itBorrowed animation
Lips closing ("m", "b", "p")0.850.430.01
Jaw0.780.670.26
Lips rounding ("w", "oo")0.680.440.02
Lower lip under the teeth ("f", "v")0.850.490.02
Corners of the mouth0.670.530.09

The borrowed column is near zero on the lips, because in a borrowed animation the lips are closing for other words. The jaw scores better than anything else there, and the jaw is why a borrowed animation looks passable from across a room.

The aligner is an extra install. Skip it and the tool guesses the timing from the silences in the audio and how long each sound usually runs, which is the middle column. The website always uses the aligner.

IN THESE GIGS // WHERE IT GETS USED

EVERY RECORDED LINE ON THIS SITE

This got written because the gigs needed it. Every character they invent is read by a real person, and for most of a year those lines ran on borrowed animations. On 27 October 2026 all three of the older gigs were rebuilt with their own, in one pass. The two newest mods have had them since their first release.

FIXER · HODAN WARSAME 1.0.0

Built with it

Hodan's lines, in her introduction call and in all six answers of her conversation, in person and on video.

GIG 04 · RETURN TICKET 1.0.0

Built with it

Hodan, Maricel, Stamps and Reynaldo, plus the lines cut short from the game's own recordings of V and Johnny.

GIG 03 · ACCEPTABLE LOSS 1.2.0

Rebuilt

No new characters in this one, so what got rebuilt is the lines cut out of Johnny's and Dino's own recordings, which still had the animation of the uncut line on them.

GIG 02 · DEAD RINGER 1.3.0

Rebuilt

Char and the merc, played by Sarah Fish and fronterawave, and the lines cut from Wakako's recordings.

GIG 01 · NEGATIVE BALANCE 2.4.0

Rebuilt

Hoshino's lines, and the lines cut from Nix's and Mama Welles's recordings.

A NOTE ON CUT LINES

Why cutting needs this too

A vanilla line used whole plays the game's own animation and needs nothing. Cut one short at a pause and the animation no longer matches what is heard, so the cut lines get rebuilt from their own audio like a recorded one.

GET IT // TWO WAYS, SAME ENGINE

THE WEBSITE OR THE TOOL

The server behind the website runs the public tool's 0.1.0 release and produces byte-identical files, so the choice is only about how much you want to install.

lipsync.tools, in the browser

No install, no game needed, no account. Upload up to 20 lines of one character, each with its recording, the exact words and the line's animation name, and get one .anims back in about half a minute. Recordings as wav, ogg or flac, up to 60 seconds each. There is a sample to try first. It makes the file and stops there. Putting it in your project and linking the scene and the lipmap is on you.

The tool, on your PC

Windows, the WolvenKit app 9.1 or later, and the game installed. The aligner is recommended and goes in through Miniforge. Point scene at a WolvenKit project and a scene and it does the whole job. An animation per line, from its recording and its subtitle. Each line linked to its animation, the scene added to the lipmap, the lipmap registered with ArchiveXL. It copies every file it is about to change into a backup folder first.

Just the animations

anims writes only the .anims files, for modders who want to link the scene and the lipmap themselves in WolvenKit rather than let the tool do it for them.

Rebuild the dictionary yourself

extract, build-dictionary and evaluate rebuild the sound dictionary from your own copy of the game and score it. The shipped one is 1.2 MB of numbers, read from game versions 2.31 and 2.5, base game and Phantom Liberty.

PLAYERS OF A MOD MADE WITH IT NEED: ArchiveXL · THE SAME AS ANY OTHER VOICED MOD · NOTHING EXTRA

TRY IT ON YOUR OWN LINES[DEMO]

THE TOOL, THE DOCS AND THE SOURCE[FORK]

NO GENERATIVE AI // ASKED OFTEN

The mouth shapes are averages of CDPR's own animations, played at the times the aligner finds. The aligner uses a trained acoustic model to locate known words and sounds in audio, and it is handed the words rather than transcribing them. Nothing comes out of a prompt. No voice is made, cloned or imitated.

Same rule as the gigs themselves: a character the base game voices speaks in the game's own recordings, and a character a mod invents is read by a person. The tool moves their mouth and nothing else.

LIMITS // WHAT IT DOESN'T DO
English only

Another language needs its own aligner model and its own dictionary built from that language's lines in the game. The contributing guide has the shape of it if you want to add one.

Only the mouth moves, plus the blinks

In CDPR's own lines the eyebrows and the direction of the eyes also move a little while a character talks. In lines made with this they don't. Everything except the mouth and the eyelids stays at the character's rest.

The words have to match the recording exactly

Numbers written out as words, "fifteen" and not "15", and names spelled as they are said. Invented names need no special handling: they get built out of their sounds like any other word. Upload the same audio file your mod ships, since a re-encoded copy can move a sound by a frame.

One set per character per scene

That is the game's rule, not the tool's. If a character speaks both new lines and the game's own lines in one scene, the tool copies the game's animations for those lines into the new set unchanged, so nothing is lost.

Tested on game 2.5, on Windows

The tool loads WolvenKit's libraries to read and write the game's files, so it runs where WolvenKit runs. The dictionary itself was read from 2.31 and 2.5.

NEXT // KEEP READING

The gigs themselves are open too. Each one is a WolvenKit project on GitHub, alongside the notes from working out how each piece of the game behaves.

BUILD YOUR OWN GIG[INJECT]

THE FIXER: HODAN WARSAME[OPEN]

THE GIGS IT SPEAKS IN[OPEN]