Audio Theater

Audio Theater

Generate multi-character audio (radio drama, lipsync voice reference, or podcast) from a script or idea. The engine is the same in all modes: one clean Gemini TTS clip per line, one voice per character. Those clean per-line clips are exactly what you need to mix a radio play, feed lipsync in seedance-2, or assemble a podcast.

Setup

Install dependencies (one-time):

pip3 install -r ~/.cursor/skills/audio-theater/scripts/requirements.txt

Keys are reused from sibling skills - you usually do not need to set anything:

• Gemini API key from the asset-generator skill (gemini_api_key).
• Replicate API token from the sound-effects skill (replicate_api_token, with the usual fallbacks).
• (recommended for realistic SFX) ElevenLabs API key (elevenlabs_api_key) — get a free key at https://elevenlabs.io and set it with setup_key.py --elevenlabs KEY. When present, the SFX backend defaults to ElevenLabs (much more realistic foley/ambience than Stable Audio). Without it, SFX fall back to the sound-effects skill (Stable Audio).

Import/verify them once:

python3 ~/.cursor/skills/audio-theater/scripts/setup_key.py            # auto-import from sibling skills
python3 ~/.cursor/skills/audio-theater/scripts/setup_key.py --show     # show what is configured

ffmpeg and ffprobe must be on PATH (brew install ffmpeg).

Modes

Music & stems. Add music cues for a score (story/podcast underscore) or scene music. When music is present the mixers emit three tracks: final.mp3 (everything), final.music.mp3 (music only), and final.nomusic.mp3 (dialogue + SFX, no music — the track to feed seedance-2 for animated storyboards, then overlay the music back). See "Three deliverables" below.

Fitting a fixed duration (reels / ads). When the audio must land at a target length (e.g. a 15s reel), budget the script to that length at the voices' natural pace — use fewer, shorter lines. Do not lean on time-stretching the speech: the Gemini voices read deliberately, so heavy atempo (>1.1×) sounds rushed and raises cognitive load — the opposite of what an ad/explainer needs. Rough budget: **2 spoken words per second** (this already accounts for the model's slow pace + inter-line pauses), so a 15s clip ≈ 25-30 words total. The model also pads each clip with leading/trailing/inter-sentence silence; trim that dead air first (it buys real seconds without changing perceived pace), and only then apply a gentle atempo (≤1.1×) if you're still slightly over. If it's still too long, cut a line — never crank the speed. Details + the trim/scale filters: REFERENCE.md → "Pacing & fixed-duration targets".

Workflow

The skill is a set of small scripts. Run only the steps the mode needs. Outputs always go to a project folder audio-theater/<slug>/ (pass --out).

SCRIPTS=~/.cursor/skills/audio-theater/scripts
OUT=audio-theater/la-tormenta        # project-relative output folder

1. Get a script (script.json)

You have a dialogue/story already, or just an idea. Either way the pipeline runs on a script.json (schema below).

• From an idea or rough draft -> let Gemini write it:

python3 $SCRIPTS/write_script.py \
  --idea "Dos marineros discuten mientras se acerca una tormenta" \
  --mode theater --language es --out $OUT
# podcast: two hosts + intro/outro
python3 $SCRIPTS/write_script.py \
  --idea "Episodio sobre por que los gatos amasan" \
  --mode podcast --language es --hosts "Lucas,Mia" --out $OUT

• From an existing dialogue -> pass a .txt (lines like Marco: La tormenta se acerca.); it is parsed and voices are auto-assigned:

python3 $SCRIPTS/write_script.py --script-file dialogo.txt --mode theater --language es --out $OUT

Either way it writes $OUT/script.json + a human-readable $OUT/story.md. Review/edit script.json (voices, tags, pauses) before generating audio.

2. Generate the voices (lines/*.wav, dialogue.wav, lines.json)

One clean clip per line, concatenated into dialogue.wav with exact timing in lines.json:

python3 $SCRIPTS/generate_voices.py --script $OUT/script.json --out $OUT
# Exactly 2 speakers and you want a single natural conversation take:
python3 $SCRIPTS/generate_voices.py --script $OUT/script.json --out $OUT --two-speaker

It warns if any clip exceeds --max-clip-seconds (15, the seedance limit). Keep lines short for lipsync.

3a. (lipsync mode) Export the seedance manifest

python3 $SCRIPTS/export_lipsync.py --out $OUT

Writes $OUT/lipsync.json (per clip: speaker, voice, exact transcript, duration, ok for <=15s, suggested panel). Then hand off to seedance-2 (see "Lipsync handoff" below). In lipsync mode you can stop here - the clean clips are the deliverable.

3b. (theater/podcast) Transcribe with word timecodes (words.json)

python3 $SCRIPTS/transcribe.py --audio $OUT/dialogue.wav --script $OUT/script.json \
  --backend auto --out $OUT

--backend: replicate (WhisperX, default for auto when a token exists), local (faster-whisper), or auto (replicate then local fallback).

4. (theater/podcast) Author the SFX/music sheet (cues.json)

You (the agent) write $OUT/cues.json by reading script.json + words.json and placing sound where it makes sense (use word timecodes to anchor one-shots). Cue type is ambient, oneshot, or music. Schema below.

5. (theater/podcast) Generate the cued sounds (sfx/*.mp3)

python3 $SCRIPTS/generate_sfx.py --cues $OUT/cues.json --out $OUT
# force a backend:  --backend elevenlabs | sound-effects   (default: auto)

ambient/oneshot SFX use the --backend (default auto):

• elevenlabs (auto-selected when an ElevenLabs key is set) — ElevenLabs Sound Effects (eleven_text_to_sound_v2). Best realism for foley/ambience; ambient cues are rendered as native seamless loops (loop=true). Send a plain natural-language description (no keyword soup); tune realism with optional cue field prompt_influence (0-1, default 0.4).
• sound-effects — Stable Audio 2.5 via the sound-effects skill (synthesis; fine for abstract/UI, weaker on realistic foley). Uses the cue category.

music cues are always instrumental (no sung vocals). Pick the backend with --music-backend (default auto → hq):

• hq — MiniMax Music 2.6 with is_instrumental=true (via audio_music.py). The model ignores lyrics and never sings; the prompt (built from description + mood) drives the whole track. Best for scores, beds, scene music.
• fast — the bg-music skill (quicker, lighter instrumental tracks).

We intentionally do not use bg-music-hq for theater music: it targets MiniMax music-2.5, which has no instrumental flag and routes structure directions through the lyrics field — so music-2.5 sings them literally ("soft piano joins…"). audio_music.py reads bg-music-hq's mood library read-only to build the style prompt, but does not modify that skill.

A per-cue music_backend overrides the run default. Pass the cue mood (e.g. pet-lullaby, cinematic, podcast-bed); audio_music.py --list-moods shows the library. The model returns a full-length track (no looping needed); if it's shorter than the window the mixer loops it.

python3 $SCRIPTS/generate_sfx.py --cues $OUT/cues.json --out $OUT --music-backend hq

Files land in $OUT/sfx/ and durations are written back into cues.json.

6. (theater/podcast) Mix (final.mp3 + stems)

python3 $SCRIPTS/mix.py --dialogue $OUT/dialogue.wav --cues $OUT/cues.json --out $OUT

One-shots are placed at their timecode; ambient beds are looped/trimmed to [start,end] with fades and sidechain ducking keyed off the dialogue. Music behaves like a background score: it sits well below the dialogue and gently lowers under the whole content (voices and SFX) — but the ducking is smooth/broadcast-style (slow release, shallow ratio) so it never pumps up and down between words and never gets louder than the voices. Keep music base levels low (-24..-20) and SFX peaks under the dialogue. Everything is summed and loudness-normalized to final.mp3.

Stems (--stems, default auto). When the project has music cues, the mixer also emits two extra MP3s next to final.mp3 (this is the standard music deliverable — see "Three deliverables" below):

• final.nomusic.mp3 — dialogue + SFX, no music (the track to feed seedance-2 for animated storyboards).
• final.music.mp3 — the music only (ducked + faded exactly as it sits in the full mix).

All three share one linear normalization gain, so final = nomusic + music exactly (you can overlay the music back after the video render). --stems always forces it even without music; --stems off writes only final.mp3.

6b. (theater, optional) Spatial stereo mix (final.mp3)

For an immersive radio-drama feel, mix_spatial.py places each voice line and each one-shot SFX on a virtual stereo stage (left/right + near/far + movement) instead of the flat center mix of mix.py:

python3 $SCRIPTS/mix_spatial.py --out $OUT --output-name final.mp3
# headphone-friendlier hard pans:  --crossfeed
# wider voices:  --voice-pan-limit 0.6     stronger distance:  --max-atten-db -14
# keep SFX full level under narration:  --no-duck-sfx

It reads per-line clips from lines.json, character seats + per-line moves from script.json, and per-cue positions from cues.json (all the new spatial/stage fields below are optional and backward compatible). With no positions authored it still seats non-narration speakers at gentle alternating L/R offsets for a natural stereo image. Ambient SFX beds stay stereo (optional balance/distance) and keep their ducking; voices and one-shots are panned/moved.

mix_spatial.py emits the same stems as mix.py (final.nomusic.mp3 + final.music.mp3 when there are music cues; --stems controls it), and music sits under the content bus with the same smooth, shallow ducking as the flat mixer (no pumping, never above the voices).

Music: fixed score by default; subtle entrances when diegetic. Background / narrator-style score music must be FIXED (no movement) — by default a music cue is a still wide stereo bed, which is what you want for "part of the story" underscore. For diegetic scene music (a music box on the table, a radio across the room) give it spatial with a single, gentle gesture — never a zig-zag (on/off/on/off reads as broken):

• {"scene": true, "enter": "left" | "right" | "front", "pan": …, "distance": …} — glides in from that side/depth over a few seconds, then holds at its seat (pair with fade_in).
• {"scene": true, "exit": "left" | "right" | "front"} — holds, then glides out at the very end (pair with fade_out).
• {"scene": true, "from": {…}, "to": {…}} — one clean sweep across the window. A multi-point path is collapsed to first→last (no bouncing).

Scene music is collapsed to a positioned point source and still ducks under the content.

Narration always wins. The narrator sits closest (smallest distance) and one-shot SFX are kept behind the dialogue plane and ducked under the narrator's voice so a loud effect (a door, a gust) never buries the narration:

• --sfx-min-distance (default 0.12): floor on how close any SFX may render, so the narrator (≈0.05) always reads as the closest source. Movement still works; SFX just can't come closer than this.
• one-shot SFX duck under the narration voice (keyed off the narrator, not the whole dialogue, so SFX stay punchy during character lines and gaps). --sfx-duck-db (default -6) sets how much they dip; --no-duck-sfx disables it; --no-duck disables all ducking. If there's no narrator, SFX duck under all voices instead.
• For an audio-theater feel, move things boldly — wide pans and big distance changes read far better than subtle ones. A sound that recedes (distance increasing) or a voice that approaches (distance decreasing) is the clearest cue; see the historia-breve demo (Lila walks in from far-left to center, the owl flies across L→R and away, the gust sweeps past and recedes).

Theater only. The lipsync feed (split_tracks.py → seedance-2) must stay centered/mono-safe for the video model, so it keeps using the flat mix.py. Use mix_spatial.py for the standalone listening deliverable, not for lipsync audio.

6c. Three deliverables when there's music (story, podcast, storyboard)

Whenever a project has music cues, both mixers deliver three tracks in one pass (the music stem is ducked + faded exactly as in the full mix, and the three share one gain so they recombine):

Storyboard / seedance workflow. Audio-driven video models get confused by music, so feed them the no-music track, then layer the music back over the rendered video:

# 1) render the storyboard with the music-free track as the audio reference
#    (use final.nomusic.mp3, trimmed to < the render duration; see "Lipsync handoff")
# 2) after the video is back, overlay the music stem on top:
ffmpeg -y -i reel.mp4 -i final.music.mp3 \
  -filter_complex "[0:a][1:a]amix=inputs=2:normalize=0:dropout_transition=0,loudnorm=I=-16:TP=-1.5:LRA=11[a]" \
  -map 0:v -map "[a]" -c:v copy -c:a aac -shortest reel_final.mp4

Because the music stem keeps the original timecodes (and the same ducking it had in the full mix), it lines up automatically and still dips under the dialogue. For projects that also split narration vs on-camera voice, combine this with split_tracks.py (below): feed lipsync_mix.mp3 (which is already music-free when you mix the no-music track) and overlay both narration.mp3 and final.music.mp3 afterward.

7. Build the transcript (transcript.md)

python3 $SCRIPTS/build_transcript.py --out $OUT

Timecoded transcript with [SFX: id] placeholders (theater/lipsync) or speaker show-notes (podcast), plus a list of the SFX/music used.

script.json schema

{
  "title": "La tormenta",
  "language": "es",
  "language_code": "es-US",
  "mode": "theater",
  "characters": [
    {"name": "Marco", "voice": "Charon", "persona": "marinero veterano, voz grave y cansada", "stage": {"pan": -0.35, "distance": 0.18}},
    {"name": "Inés", "voice": "Aoede", "persona": "joven grumete, nerviosa", "stage": {"pan": 0.4}},
    {"name": "Narrador", "voice": "Sulafat", "role": "narration", "on_camera": false, "persona": "voz cálida que cuenta la historia"}
  ],
  "lines": [
    {"index": 0, "speaker": "Marco", "text": "La tormenta se acerca, muchacha.", "tags": ["serious", "tired"], "pause_after": 0.4},
    {"index": 1, "speaker": "Inés", "text": "[trembling] ¿Llegaremos a puerto?", "tags": ["panicked"], "pause_after": 0.2,
     "spatial": {"from": {"pan": 0.6, "distance": 0.6}, "to": {"pan": 0.2, "distance": 0.15}}}
  ]
}

• language / language_code: language is the 2-letter content language (es). language_code is the BCP-47 accent/locale Gemini TTS renders in — set it to control the accent: es-US (LATAM-neutral, the default for es), es-ES (Castilian), pt-BR, en-US, etc. The 30 voices are accent-flexible; the locale, not the voice, drives the accent. Resolution order: --language-code flag > script language_code > mapped default from language (es→es-US). Omit for auto-detect.
• voice: one of the 30 Gemini voices (see scripts/voices.json). Auto-assigned if omitted. Timbre only — pair with language_code for the accent (e.g. for neutral LATAM Spanish keep language_code: "es-US").
• tags: optional inline delivery tags (e.g. whispers, laughs, excited). They are woven into the line as [tag] audio tags. You can also put tags directly inside text.
• pause_after: seconds of silence appended after the line in dialogue.wav (default 0.3).
• role / on_camera: optional. Mark a character as off-camera narration with "role": "narration" (also accepts voiceover/vo/offscreen) or "on_camera": false. Only used by split_tracks.py to separate the narrator's voice from the lip-synced characters (see "Narration vs on-camera" below). Default: on-camera.
• stage (character, spatial mix only): the character's default seat {pan, distance} — pan ∈ [-1,+1] (L..R, 0=center), distance ∈ [0,1] (0=close, 1=far). Used by mix_spatial.py; ignored by mix.py. Default: narration centered/intimate, others auto-seated L/R.
• spatial (line, spatial mix only): per-line override of the seat — static {pan, distance}, or movement {from:{…}, to:{…}} / {path:[{t,pan,distance}, …]} (t in seconds from the line start). Voice pan is clamped to ±--voice-pan-limit (default 0.5) for intelligibility.

cues.json schema

{"cues": [
  {"id": "rain_bed", "type": "ambient", "description": "lluvia constante sobre cubierta de madera",
   "category": "nature", "start": 0.0, "end": 42.0, "gain_db": -18, "duck_db": -8, "fade_in": 1.5, "fade_out": 2.0},
  {"id": "door_slam", "type": "oneshot", "description": "portazo de madera pesada",
   "category": "foley", "start": 12.6, "gain_db": -4, "spatial": {"pan": 0.0, "distance": 0.1}},
  {"id": "owl", "type": "oneshot", "description": "un búho ululando mientras cruza el bosque",
   "start": 19.5, "gen_seconds": 3, "gain_db": -6,
   "spatial": {"path": [{"t": 0, "pan": -0.9, "distance": 0.85}, {"t": 3, "pan": 0.9, "distance": 0.85}]}},
  {"id": "story_score", "type": "music", "mood": "cinematic", "music_backend": "hq",
   "description": "soft mystical underscore, slow strings and distant music box, low and unobtrusive",
   "start": 0.0, "end": 120.0, "gain_db": -18, "duck_db": -14, "fade_in": 3.0, "fade_out": 4.0},
  {"id": "music_box", "type": "music", "mood": "pet-lullaby", "description": "tiny wind-up music box on a table",
   "start": 60.0, "end": 80.0, "gain_db": -12, "spatial": {"scene": true, "enter": "left", "pan": 0.5, "distance": 0.35}},
  {"id": "intro_music", "type": "music", "mood": "podcast-intro", "description": "cortina alegre de podcast de mascotas",
   "start": 0.0, "end": 8.0, "gain_db": -10, "duck_db": -12, "fade_out": 2.0}
]}

• type: ambient (looping bed, ducked under voice), oneshot (single hit at start), music (instrumental score/bed/jingle).
• description: natural-language prompt. With the elevenlabs backend, send it plain (the model handles natural language); name concrete sources/materials, one source per cue. For repeated/continuous actions describe the full sequence, not a single hit — e.g. "a sequence of several footsteps crunching on leaves", "several twigs snapping in succession" — and set a matching gen_seconds (~3-5s). Otherwise the model renders one isolated event (one step, one snap), which sounds thin.
• For a stylized / dramatized feel, push the description (e.g. "mystical eerie wind with an otherworldly hum" instead of plain realistic wind) and lower prompt_influence (~0.3) so the model adds character.
• prompt_influence (ambient/oneshot, elevenlabs only): 0-1, default 0.4. Lower = more natural variation / more stylized, higher = follows the prompt more strictly.
• category (ambient/oneshot, sound-effects backend only): nature, ambient, foley, mechanical, transition, game, voice, generic. Ignored by the elevenlabs backend.
• mood (music): a mood from the bg-music-hq library (read-only) used to steer the prompt — podcast-intro, podcast-bed, podcast-outro, cinematic, ambient, pet-lullaby, etc. (run audio_music.py --list-moods for the full list).
• music_backend (music only): hq (MiniMax Music 2.6, is_instrumental=true — default; never sings) or fast (bg-music, quicker). Overrides the run's --music-backend. Music is always instrumental.
• gain_db: level offset applied to the cue. Dialogue is always on top — SFX and music must sit under it. SFX should be present but not louder than the voices: one-shots usually -12..-8 (peaky sources like a door/gust/snap lower, ~`-13..-11; quiet textures can go higher), SFX beds -18..-15. **Music score** sits lowest — base -24..-20; it stays a soft bed and the gentle ducking only nudges it a few dB further down under content. duck_db: how much the cue dips while content plays — -4..-8for everything (music's duck is intentionally **shallow + smooth** so it doesn't pump; default music-8, capped to a gentle ratio with a long release). start/end/fade_in/fade_outin seconds.gen_seconds` (oneshot) pins the generated length.
• spatial (spatial mix only, mix_spatial.py): position on the virtual stage. One-shots are panned/moved like voices — static {pan, distance}, or movement {from, to} / {path:[{t,pan,distance}, …]} (SFX pan clamped to ±--sfx-pan-limit, default 0.95). Ambient beds and score music stay stereo & FIXED; they only honor an optional {pan} (gentle L/R balance, keeps width) and/or {distance} (level + low-pass). A music cue becomes a positioned point source (diegetic scene music) only when its spatial has {"scene": true}; then keep the motion subtle and single-direction — enter/exit ("left"|"right"|"front") for a gentle glide in/out, or one from/to sweep. Never zig-zag music. Ignored by mix.py.

Lipsync handoff to seedance-2

The clean per-line clips are the audio reference for seedance-2 lip-sync. This skill produces the clips + lipsync.json; it does not generate the storyboard (use gpt-image-2 / seedance-2 for that). Per clip:

1. Upload the clip: Higgsfield media_upload -> PUT bytes -> media_confirm (type: "audio"). Seedance accepts audio only via medias[] and requires at least one image/video reference, with total duration <=15s (export_lipsync.py validates this with ok).
2. Build the prompt with the exact transcript:

python3 ~/.cursor/skills/seedance-2/scripts/prompt_tools.py reframe \
  "[Image1] the sailor speaks to camera" --images 1 --audios 1 \
  --audio-transcript "La tormenta se acerca, muchacha."

3. Call generate_video (model seedance_2_0) with the storyboard panel image (reference-image role) + the confirmed audio media_id (audio role). Map lipsync.json[*].panel to the storyboard panel.

Keep the audio strictly under the render duration. Seedance fails/derails when the audio length equals (or exceeds) the requested video duration. For a 15s render, feed audio ≤ 14.9s (trim with a tiny tail fade). The render comes back ~15s.

Narration vs on-camera: split tracks for video lip-sync

When a single mixed track contains both an off-camera narrator and an on-camera character that should lip-sync, feeding the whole thing to an audio-driven video model is wrong: the model tries to make an on-screen character mouth the narrator's words. Split the project into two role-separated tracks and recombine in post:

python3 $SCRIPTS/split_tracks.py --out $OUT
# or override role detection:
python3 $SCRIPTS/split_tracks.py --out $OUT --narration "Narrador" --onscreen "Doki"

It reads lines.json + cues.json (and character role/on_camera from script.json) and writes:

• narration.mp3 — only the narration/off-camera lines, on the original timeline (silence elsewhere).
• lipsync_mix.mp3 — the on-camera voices + all SFX, with the narration muted. This is the track to feed the video model: it lip-syncs only the on-camera character(s) and never the narrator.
• tracks.json — which speakers/lines went to which track + durations.

Then:

1. Trim lipsync_mix.mp3 to under the render duration (≤14.9s for a 15s reel) and feed it to seedance-2 as the audio reference (prompt: lip-sync the on-camera character only when its voice is present; mouths closed during silence/narration; do not add captions and keep panel text unchanged).
2. After the video renders, merge narration.mp3 back on top of the rendered video's audio (it already carries the on-camera voice + SFX), e.g.:

ffmpeg -y -i reel.mp4 -i narration.mp3 \
  -filter_complex "[0:a][1:a]amix=inputs=2:normalize=0:dropout_transition=0,loudnorm=I=-16:TP=-1.5:LRA=11[a]" \
  -map 0:v -map "[a]" -c:v copy -c:a aac -shortest reel_final.mp4

Because narration.mp3 keeps the original timecodes, it lines up with the visuals automatically.

Notes & limits

• TTS model defaults to gemini-3.1-flash-tts-preview. Output is PCM 24kHz 16-bit mono, wrapped to WAV.
• Gemini multi-speaker is capped at 2 voices and quality drifts past a few minutes - that is why the default is one clip per line. --two-speaker exists for exactly-two-speaker scenes/podcasts.
• The model occasionally returns text instead of audio (random ~500); generate_voices.py retries automatically.
• Keep lines short, especially for lipsync (one clip per shot, <=15s).
• Hitting a fixed total length is a scripting job, not a speed job — write fewer/shorter lines (~2 words/sec budget) and trim dead air; reserve atempo for a gentle (≤1.1×) final nudge. See "Fitting a fixed duration" above.

Additional resources

• Voices, tags, WhisperX/cues/lipsync schemas, mix internals, podcast recipes, troubleshooting: REFERENCE.md