> ## Documentation Index > Fetch the complete documentation index at: https://docs.seriesmith.com/llms.txt > Use this file to discover all available pages before exploring further. # Recorded dialogue: lip-synced episodes > Lip-synced episodes from real recorded lines instead of generated voices. Use this when you have real recorded lines (actors, you, or TTS) instead of letting H3 clone or invent voices. ## Steps ``` cd pip install faster-whisper # once seriesmith align Shows\ep05 audio\ep05_dialogue.wav --dry-run # look first seriesmith align Shows\ep05 audio\ep05_dialogue.wav seriesmith build Shows\ep05 seriesmith render Shows\ep05 --proxy seriesmith assemble Shows\ep05 --proxy --audio master # picture + your recording, to check sync ``` 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 `\_history\`. * **Series config:** sets `audio.mode` to `source_track` and `audio.track` to your recording. The old copy is kept in `\_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 `.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`. | Retention | What H3 does with your line | Default for | | - | - | - | | `fully_copy` | Your clip becomes the shot's entire soundtrack. Lips follow it, and no sound effects are added. | `dub` | | `partially_copy` | Copies your line exactly and adds ambience and action sounds around it | `dub_keep_foley` | | `reference` | Only borrows the voice and timing, as before. Lip sync is loose. | | 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: | Flag | What it does | | - | - | | `--dry-run` | Report only. Changes no files. Always worth running first | | `--retranscribe` | Ignore the cached `.words.json` after replacing the recording | | `--words ` | Use word timings you already have instead of transcribing | | `--model ` | Whisper model: `tiny.en`, `base.en`, `small.en`, `medium.en` (default), `large-v3` | | `--device cuda\|cpu` | Where Whisper runs. `auto` by default | | `--script ` | Pick the script when the folder holds more than one `.md` | | `--no-snap` | Leave windows exactly as measured instead of lengthening them to the target's frame grid | | `--policy`, `--retention` | Override the defaults in the table above for the whole episode | ## 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. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.