Skip to main content
Everything you hear in an episode besides the picture, made and mixed in seriesmith: the dialogue, a bed of room tone for each place, the shots’ effects, and a score written to the cut. This page is the order of work and what each part does. Scoring itself, cue by cue, is in SCORING.md; every button is in EDITOR.md (“Sound: the episode’s mix”), and every command in COMMANDS.md.

​
The idea in one paragraph

A model like H3 renders each shot with its own sound: the lines, footsteps, a room tone and often a bit of music, different in every take. Cut together, that sound changes at every cut. So the episode’s sound is rebuilt on four tracks under the picture. Dialogue keeps only the voices from each take (the rest split off). Ambience lays one bed of room tone under every shot of a place, so the room carries across its cuts. Effects adds a shot’s sync sound, made from its frames, where it needs one. Music is the score, drafted from the script’s story beats and rendered cue by cue. The mix sets the levels and dips the music and ambience under the lines, and the export writes the finished sound, its stems and an M&E.

​
Before you start

  • Lock the picture, mostly. Everything on the sound tracks is anchored to a shot and a frame of its take, so it follows trims and reorders. But a cue is written to the cut’s timing, and re-cutting a scene means making its cues again.
  • Every shot needs a take in the pass you’re working on. A cue, a bed and a shot’s effects are all timed to the rendered shots. The proxy and final passes each have their own mix: score the pass you’ll deliver, or the proxy first and the final once it’s rendered.
  • Install what you’ll use. The editor’s Features window lists each part under Sound and downloads it: the voice stems (Hybrid Demucs, seriesmith’s own node), the ambience model (MMAudio, the default), the effects models (H3 FL2VA, the default) and the music models (Stable Audio 3, the default; YuE2 for songs). What each needs is in FULL_SETUP.md, “Sound”.
  • Open the sound lanes. Sound in the timeline’s header unfolds a hits lane and the four tracks under the clips: Dialogue, Effects, Ambience, Music. Scoring, beside it, opens the spotting sheet.

​
The order of work

  1. Voices only: the dialogue track keeps each take’s voices, nothing else.
  2. Ambience beds: a bed for each place, laid under its shots.
  3. Effects: for the shots whose own sound is gone or wrong.
  4. The score: a first pass drafted from the script, then each cue shaped and rendered.
  5. The mix: levels, ducking, a place where the music should carry over the lines.
  6. Export: the episode with its mix, the stems and the M&E.
Steps 1 to 3 change what’s under the music, so do them first: the score is judged against the sound it will sit on.

​
1. The dialogue: voices only

Each take’s own sound has its lines and everything else the model added, its music and room tone. Voices only splits each take into its voices and the rest (Hybrid Demucs, on your ComfyUI) and plays only the voices, so the beds, effects and score aren’t fighting a different bit of music at every cut.
  • Turn it on: right-click the Dialogue lane ▸ Voices only, or its head’s ⋮, or the Refs tab’s Dialogue section.
  • Make voice stems beside it splits every take the cut plays: a job a take in Jobs, about a second each once loaded. It says how many are split (“8/10 shots split”). A take rendered again is split again when you ask.
  • A shot whose stem isn’t there yet plays its whole sound, never silence, with a warning mark on the lane. A shot the script gives no lines is silent on the Dialogue track (its stem would only hold the model’s ad-libs and murmurs).
  • A shot can go its own way: its clip’s Sound menu ▸ Keep its whole sound (a take whose own music you want), Voices only (a gasp the model gave a shot with no lines), or Split its voice again.
  • Command line: seriesmith voices <episode> (--proxy, --shots, --redo, --list).
With the takes’ own room tone and effects gone, the next two steps put them back, once and consistently.

​
2. Ambience: a bed for each place

A place’s ambience bed is its room tone or air (a refrigerator’s hum and a neon sign’s buzz, rain on a window, wind and gulls) made once from words and looped under every shot set there, crossfaded into the next place’s at the cut. It’s what makes a room sound like one room across its cuts.
  • Say what’s heard there. A location’s ambience in the series config, written as a sound designer would: the steady bed first, then a few sparse events, never music or voices (AUTHORING.md, “What a place sounds like”). Or write it in the editor (below).
  • Make it: the Refs tab’s Ambience section lists each place, its words, how many shots it plays under and whether its bed is made: open a row to Make it, hear the takes and pick one. + Bed on the Ambience head lists the places the cut is at, takes new words (written into the series config) and makes the bed. Or right-click the Ambience lane over a shot ▸ Make <place>‘s ambience bed.
  • On the lane the beds are dashed blocks named for their place, laid from the cut (they follow trims and can’t be dragged). Right-click one for its level or Mute this place. An angle or a relit version of a place plays its place’s bed.
  • A sound file dropped on the Ambience lane plays over the beds (a walla track, a passing train).

​
3. Effects: a shot’s sync sound

Effects are a shot’s footsteps, the door, the cups on the table, made from the frames of the take the cut plays. Many takes already have good effects in their own sound; with voices only on, those are gone too, so make effects for the shots where you miss them.
  • Give the shot a sound: line first (“cup taps on formica, hands clapping”): the effects are made from it, else from the action, and an action written for the picture can mislead (“…looks up, about to sing” gets a voice).
  • Make them: a clip’s Sound ▸ Make effects (several clips selected: one take each), or + Effects on the Effects head (the cut’s shots by sequence, ticked where there are none), on MiniMax H3 (the default, about 10 s a shot) or, under Make effects on…, LTX-2.3 Foley or HunyuanVideo-Foley XL (a minute or two). Each new take is picked into the mix as it lands.
  • On the lane a shot’s effects sit at its clip. Drag an edge to trim them (a voice creeping in at the end), drag the corner squares for fades, double-click the line for a volume point; right-click for its level, its other takes (▶ to hear one, Pick) and Make again.
  • When the cut plays another take of the shot, its effects are stale: dashed, left out of the mix until they’re made again, or Keep as it is.
  • Command line: seriesmith fx <episode> --shots sh010,sh020 (--target, --takes, --pick, --list).

​
4. The score: from the script to cues

The music is made of cues: pieces of score over a stretch of shots, each rendered as several takes, measured and ranked, one picked into the mix. You can place every cue by hand (SCORING.md step by step), but the quickest start is a first pass drafted from the script.

​
Mark the story’s beats

The draft reads the script’s story structure, written as comments (they change nothing the video models see, so nothing renders again):
  • // structure: names one of the structures in STORY_STRUCTURES.md (Pixar Story Spine, Sitcom, Fable, Episodic Drama, the Hero’s Journey…), and // beat: sits above the first shot of each beat, with the beat’s name as that structure writes it. The script-writing skill writes them for a new script.
  • An older script without them: the Draft dialog says so, and its Structure list lays the beats out for you (each where it falls in a story, on the nearest sequence) and can write them into the script, to move by hand where a beat really starts. Or set a shot’s Beat in its form (More).
  • // theme: show, tender above a shot puts a theme on a moment the beats can’t find (the heart of a comedy, a look that changes the scene): held low to the end of its scene, or a sting when the words say sting. // flashback above a sequence makes its music an echo, thinner and further away.

​
Draft

Scoring ▸ the Spotting tab ▸ Draft… (or seriesmith score-draft <episode>, --write to write it). It shows the cues it would make, and writes nothing until you press Write.
  • Each structure has a Music table: each beat scored, held low under the lines, a short sting, or played dry, with its mood, its energy and where the music turns, peaks or drops.
  • Beats in a row with music are one cue, a section each, so the score changes with the story rather than stopping at every beat. A story told straight through (a bedtime story on the Story Spine) is one cue across the episode; a sitcom plays its scenes dry with short stings and a bumper of the show’s theme at each change of place.
  • A show whose episodes share a shape of their own describes it once in the series config, sound.format (AUTHORING.md): its beats’ names and what each gets, and gaps for a talky beat (music only where no one speaks).
  • Write adds the draft cues (d01, d02…) and their hits, and turns the Dialogue track to voices only. Nothing is rendered yet. A cue of your own is never touched; drafting again replaces only the draft cues you haven’t rendered (or, ticked, the rendered ones too).

​
Shape, render and pick

Open a cue (double-click it on the spotting map; the cue in the window’s head picks another): the Score tab shows its sections, hits, arc and takes on one bar grid. Adjust the words and the energy, Render its takes (eight, a minute or so on a 5090 for Stable Audio 3), listen under the picture (▶ beside a take on the spotting sheet) and Pick one: it lands on the Music track, its level shaped to the arc. SCORING.md has every step, fixing part of a take and the levels.

​
Themes across a show

A theme is a tune that comes back: the show’s own, a character’s, the thread of a serial. Here a theme is a picked take: render a cue with no theme, pick the take you like, and on its row press ★ Series theme: a name, its role (show, story, runner), whose it is (a character’s theme) and what it is in words. It’s written into every episode’s series config.
  • Any cue’s Theme can then be that series theme: Stable Audio 3 renders the cue from the theme’s take (at a strength, 0.86 by default), so the tune recurs in the cue’s own instruments, fitted to its tempo and bars.
  • With SheetSage2 installed, each take gets a melody score (how much of the theme it kept); takes that lost it rank last.
  • The draft uses them: the Music tables name a beat’s theme by role, story (the episode’s own, made by its first scored cue), show and runner. A character’s theme (whose set) is the episode’s own whenever that character leads it, carried from episode to episode. Make the themes in the first episode, then draft the others with them.

​
Hits

A hit is a moment the music must answer: a door slams (accent), the music peaks here (arrive), cuts out (drop) or changes (turn). The draft places turns at each beat; add your own on the hits lane (double-click), at Play all’s playhead (+ Hit at playhead on the spotting sheet), or in a shot’s form (More ▸ Hit, on its first frame). Drag a hit on the hits lane or the spotting map to move it. A cue’s tempo is fitted so its hits fall on bar lines.

​
5. The mix

  • Levels: each lane’s fader (dB), M mute and S solo. Music starts at -12 dB, so the score sits under the lines. The series config’s sound.tracks sets a show’s defaults.
  • Ducking: music and ambience dip under speech (music -9 dB, ambience -6 dB), worked out from the dialogue track, the same in Play all and the export.
  • Duck regions: right-click a lane ▸ Duck region here for a stretch where the music carries the scene over the lines (score forward), or dips deeper under busy dialogue.
  • Items: drop any sound file on a lane, drag it, trim it, fade it, shape it with volume points; Plays in the mix unticked keeps it but silent. More tracks of your own (a song under the score, a walla track) from a lane head’s ⋮.
  • Play all plays the mix as the export will. Every change is saved at once and undone with Ctrl+Z.

​
6. Export

Export ▾ ▸ Export (or seriesmith assemble) renders the mix into the episode: every track at its level and ducking, brought to the loudness target (the series config’s sound.loudness, else -16 LUFS and -1 dBTP), with the stems beside it (<cut>_stems/: one per track, me for music and effects without the dialogue, and mix) and a loudness report. Export ▾ ▸ Premiere / Resolve XML lays the same on separate tracks for an editor to finish. See COMMANDS.md, “assemble”, and FINISHING.md for the loudness.

​
From the command line

Ambience beds and cue takes are made from the editor (the Refs tab, the cue window), or by an LLM over seriesmith mcp, whose score_draft tool drafts the score.

​
Quick reference