Skip to content

Character pieces

github-actions[bot] edited this page Nov 21, 2026 · 3 revisions

The Udder Truth's host is performed to playback: a script read in her voice, the performer lip-syncing to it on camera, a video model turning the take into her.

clippa narrate speak script.txt                       # the read, in [narrate] voice_id
# film yourself lip-syncing while the read plays out loud (one or more takes)
clippa perform prepare take.mov --read script.read.wav --take 1   # cut at the read's pauses
# render each pieceN.mp4 with the character model -> renderN.mp4 (Runway, WAN 3 Prime)
clippa perform stitch take.perform                     # final.mp4 + final_phone.mp4

clippa narrate voice <take> re-voices a take in her voice when the performer's own delivery is wanted instead. Setup is below.

Narration voice (ElevenLabs)

clippa narrate voice <take> re-voices a performance (a phone take, a webcam clip, an audio file) with the ElevenLabs Voice Changer: the words, timing and delivery stay the performer's, the voice becomes the character's. A video comes back as <take>.voiced.mp4 with the picture untouched. The performer's accent carries through.

  1. Make an ElevenLabs account (a paid tier for commercial use). Reads and voiced takes come back as lossless 48 kHz WAV ([narrate] output_format = "pcm_48000", Creator plan and up); a plan without it gets a 128 kbps MP3 instead, with a note. Profile → API keys → create a key (it needs the Voice Changer and Voices permissions). Paste it, alone on one line, into secrets/elevenlabs_api_key.txt (gitignored), or set the Windows user environment variable ELEVENLABS_API_KEY, which wins if both are set.
  2. Pick the voice: clippa narrate voices --accent british --gender female --age old lists your voices and matching library voices. Add a library voice with clippa narrate voices --add OWNER:VOICE_ID --name "Mother Moo".
  3. Put the printed id in clipper.toml:
    [narrate]
    voice_id = "..."
  4. clippa doctor checks the key, the plan and the voice.

Takes longer than [narrate] max_s (300 s) are split at pauses first.

clippa narrate speak script.txt reads a script in the same voice (text to speech, [narrate] tts_model = eleven_v3; inline cues like [sighs], [pause], [whispers] steer the delivery) and writes script.read.wav. The performer can then lip-sync to it on camera, so the face and timing are theirs and the voice and accent are the read's.

Clone this wiki locally