Watch tutorial 11 on YouTube
Voices and your own dialogue: who says the lines, and how they sound.
audio.mode):
This continues on the Dean episode (four lines between Dean and Pip).
1. A voice sample
1
Describe the voice
A character’s card (Refs, or Series settings’ cast) has Voice, in words: a bright,
gentle little boy’s voice. With generate, that’s all the model gets, so the voice is
made up anew each shot.
2
Generate a sample
Refs → the character’s Voice sample row → Generate a sample: a few seconds of that
voice, made on LTX-2.5 (audio only). Generate a few and listen.
3
Or import one
Drop a recording of the voice you want on the slot (or Upload…), or Use a line from a
take… to take it from a take you like.
4
Pick it
Pick the one to use. It’s written to
refs/voices/ and its voice_sample line added to
the series config.Refs’ Generate missing makes the pictures, not the voices: generate each voice from its
own row. Generated samples need LTX-2.5’s files (Settings → Features…, Voices).
2. Cloned voices
1
A sample for everyone who speaks
Each speaking character needs a picked voice sample (part 1).
2
Turn cloning on
For the whole series: Series settings → Dialogue → clone, and Save. For one
shot:
policy: clone on its lines in the script.3
Re-render
The dialogue shots are now stale: Re-render stale. Same seeds, so only the voice
differs.
4
Compare
Compare with… each shot’s old and new take: before, a voice made up for that shot;
after, cloned from the sample.
3. Your own recording
1
Install faster-whisper (once)
Align transcribes the recording with Whisper. Install it into the Python that runs
seriesmith:(Inside ComfyUI, that’s ComfyUI’s Python; for the page of its own, the repo’s
.venv.)2
Record the lines
All the episode’s lines, in script order, in one file, with a pause between each and
real pauses where the script has silent shots. Dry: no music under them. A phone is fine.
For this episode:
Are you lost, little owl? · I can’t see my tree. · Follow the stars, Pip! · Goodnight, Pip.
3
Attach it
Open Recording (the speaker icon in the Shots toolbar). Under Attach a recording,
browse to the file (Browse the ComfyUI machine…) or type its path and Use this
path. It’s copied into the episode’s
audio/, and the episode now dubs to it.4
Dry run, then Align
Dry run shows what align would write without writing it. Then Align: it transcribes
the recording, finds each line’s window, and matches it to its shot.
5
Read the report
Every line, its shot and its window. Align writes
audio: in-out on each shot (as long as
you took to say it), sets the series config’s dialogue to the recorded track, leaves
align_report.md beside the episode, and rebuilds. The old files go to _history/.6
Re-render
Re-render stale: each dialogue shot lip-syncs to your line, keeping the model’s own
sound effects under it (the default,
dub_keep_foley).7
Play it with your voice
Play all → recording: your voice under the cut.
Next
12 · Editing a real episode
Fixing a long episode without watching it all again.
Recorded dialogue
Align’s options, the audio policies, and the commands.