Skip to main content
Use this when you have real recorded lines (actors, you, or TTS) instead of letting H3 clone or invent voices.

​
Steps

Put the recording inside the episode folder (for example ep05\audio\) so the series config can point at it with a relative path. faster-whisper isn’t one of seriesmith’s dependencies: install it into the Python that runs seriesmith. The editor does the same from its dialogue track: attach the recording, then align (seriesmith/track.py).

​
Recording

  • Record the lines in script order, dry: no music under them. If there is music, split it first with python -m demucs --two-stems vocals file.wav and use the vocals stem.
  • Leave real pauses where the script has silent shots. align stretches or squeezes those shots to fit the pause, and warns you when the pause is far off.
  • One file for the whole episode.

​
What align changes

  • Script: adds audio: in-out to every shot and turns old dur: lines into comments. The old copy is kept in <episode>\_history\.
  • Series config: sets audio.mode to source_track and audio.track to your recording. The old copy is kept in <episode>\_history\ too (the same history the editor writes, so both show the same versions).
  • Report: writes align_report.md with every window and line, plus any problems it found.
  • Transcript cache: saves the transcript as <recording>.words.json, so re-running is instant. Use --retranscribe after you replace the recording.
  • Cut points: cuts land at the quietest point of each pause. Speaking shots are lengthened to the frame grid of the target they render on (H3’s 17k+5, LTX’s 8k+1) when the pause has room.

​
Audio policy and retention

A speaking shot defaults to dub_keep_foley: H3 lip-syncs to your line, and the saver keeps H3’s sound effects with the voice removed, as _foley.wav. To override, set it per shot (retention: fully_copy), for the episode ("audio": {"retention": ...}), or for the default policy ("audio": {"default_policy": "dub"}). You can also pass --policy or --retention to align. Shots in clone and generate mode are unchanged. Frame lip-sync shots as close or medium shots. The reference image’s framing matters more than the prompt.

​
align options

python -m seriesmith.align --help lists them all. The ones you reach for:

​
Conform

  • Trimming: assemble trims each timed shot back to its exact window, so the cut lines up with the recording from the first window’s start. --no-trim turns this off.
  • Sync check: --audio master lays the recording under the cut. It’s a check, not a mix.
  • Final mix: do it in Resolve. Put the cut on one track and the recording on another, starting at the time shown in align_report.md. Add each shot’s _foley.wav at its start time (listed in _shots.txt), then music.