A silent style skill skips this page. A spoken video uses four steps, in this order.
Cleanup#
DeepFilterNet is the noise reducer the studio kit already names. The worker runs it on a recorded voice before measuring loudness. It does not run on a generated voice unless the file measures noisy.
Voice#
The worker needs a WAV and word times.
| Source | When |
|---|---|
| Operator upload | The person already recorded |
| Local TTS | No vendor key, or the brief forbids a hosted voice |
| ElevenLabs or another hosted voice | The brief names that vendor and a key is present |
Hosted voices are adapters. The core job does not require one.
Word times#
Whisper (or a faster descendant such as faster-whisper) writes word times. Captions and scene changes anchor to those times, not to a guessed duration.
The transcription skill in .agents/skills/ trims leading silence and passes proper names from the script into the prompt so names survive.
Music and effects#
The mix keeps the voice in front. Music ducks while the voice is active. Sound effects sit on speech anchors, not on a free timeline. Pixabay and other stock beds are allowed only when the sidecar stores the license. A job with no license record ships without that bed.