The studio desk at night; the right monitor shows dub's stage-light home screen, and the Kaotica Eyeball mic sits in front of it.
The room from the homepage, with the mic and the right monitor, dub's home screen, lit.

04 · dub · party game

Jackbox, but you voice the scene.

A party game for friends: a movie, TV or game scene plays with a voice missing, you record that voice, and the room votes on the best take.

1/7

dubdub · a dubbing party game, live in your browser

The packs existed. Multiplayer didn’t.

a pack is a movie scene, a TV scene or a game cutscene with a voice taken out, for you to fill in. The Choicer Voicer (someone else's tool) let you record one alone; dub reads the very same packs and adds a room full of friends.

A party game, not a recording booth.

Deal the roles at random, or everyone picks their own. Grab one character or five. Nobody hears a take until the room votes.

deal characters at random or pick your own, take as many as you want, then perform blind. Live at dub.antyanty.me

Browse GameBanana. Nothing to install.

DEMO 4.1 · the pack explorerno installs

GameBanana is the mod-sharing site where the packs live. Search it without leaving the game, and dub’s own server fetches and installs the one you pick

Record, wait, and watch yourself dubbed.

press play on a pack, record when it’s your turn, then wait (solo or with friends) while the server puts your voice into the scene

Vote your favorite. Then run it back.

a couple of buttons and a good take is all it takes. The room votes on who committed hardest, then the host hits Play it again on the same scene

Four browsers, one scene, one reveal.

The only way to prove the mixer adds nothing.

CODE 4.2 · dub/tests/test_render.py:37-46dub/tests/test_render.py:37-46
def test_render_without_takes_reproduces_the_original_audio(tmp_path):
    client = TestClient(app)
    r = client.post("/api/render", data={"pack_id": "diner-argument"})
    assert r.status_code == 200, r.text
    (tmp_path / "out.mp4").write_bytes(r.content)

    to_wav(tmp_path / "out.mp4", tmp_path / "out.wav")
    to_wav(FIXTURE / "video.mp4", tmp_path / "orig.wav")
    peak, _ = peak_rms_db(str(tmp_path / "out.wav"), str(tmp_path / "orig.wav"))
    assert peak < -90, f"render colored the audio: residual peak {peak} dBFS"

With no takes, the mix is the original audio. It has to come back identical, not merely close, and this is what caught the limiter’s lookahead.

Packs don't have to come from GameBanana. Here's how one gets made.

The explorer up top plays packs other people already built. This is the other half: turning any clip into one, offline, before anyone joins a room. Seven cached stages, click one to read what it does.

DEMO 4.1 · the pack builder, stage by stageclick a stage

Separation voice from backing

Splits the clip’s audio into voice and backing, so the gated backing stage later has a clean bed to duck under each line.

The pack builder: a video clip goes through separation, voice activity detection, diarization, speech recognition, line assembly and gated backing into a pack, with a human review step for the speaker guesses. It all runs offline; the game server only reads finished packs.packbuilder/ · offline, on the Mac, may import torchseparationvoice from backingVADwhere speech isdiarizationwho is speakingASRwhat they saidline assemblyone line per turngated backingthe bed, voice-freepackvideo + lines + JSONhuman review: fix the speakersserver/ reads packsno ML, CPU box
FIG 4.3 Seven cached stages and one human step. The server on the right only ever reads finished packs. From the dub README.
CODE 4.1 · dub/tests/test_server_no_ml.py:11-32dub/tests/test_server_no_ml.py:11-32
FORBIDDEN = {"torch", "torchaudio", "transformers", "whisper", "faster_whisper",
             "pyannote", "demucs", "audio_separator", "onnxruntime", "librosa",
             "soundfile", "scipy", "numpy"}


def imported_names(path):
    tree = ast.parse(path.read_text())
    for node in ast.walk(tree):
        if isinstance(node, ast.Import):
            for a in node.names:
                yield a.name.split(".")[0]
        elif isinstance(node, ast.ImportFrom) and node.module and node.level == 0:
            yield node.module.split(".")[0]


def test_server_source_imports_nothing_heavy():
    offenders = []
    for f in sorted((ROOT / "server").rglob("*.py")):
        for name in imported_names(f):
            if name in FORBIDDEN:
                offenders.append(f"{f.relative_to(ROOT)} imports {name}")
    assert not offenders, "server/ must stay ML-free: " + "; ".join(offenders)

The one architectural rule that is expensive to fix later is a test, not a hope.

Counted, with the date they were counted.

test functions
362as of 22 Sep 2026
lines of Python, 4,667 of them tests (wc -l)
10,674as of 22 Sep 2026
lines in web/, vanilla JS, CSS and HTML
5,025as of 22 Sep 2026
web dependencies: no npm, no CDN
0as of 22 Sep 2026
to build a pack from an 86 s clip, M4 Air
~130 sas of README
lines in the pack-format spec
187as of 22 Sep 2026
players in the first playtest
4as of 19 Aug 2026

Milestones 0 to 4 done. The last one needs a human.

What I don’t claim:

Python on the server, nothing to install in the browser.

Proves real-time multiplayer over WebSockets, a clean split between an ML pipeline and a CPU-only server, audio correctness tested by null test (résumé → dub).