ElevenLabs covers five audio lanes on Scenario: speech, music, sound effects, dubbing, and voice transforms. Eleven members ship side by side at authoring time, one job each: the deciding move is picking the right member, not coaxing one model into a mode. Discover them with search and treat model_schema_get as the contract: the members agree on almost nothing.
Connection and the core loop: see the scenario skill in this repo; model-agnostic audio work: the scenario-audio skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.
| Job | \nMember | \nDeciding inputs | \n
|---|---|---|
| Expressive speech, dialog | \nEleven v3 | \ntext with inline tags | \n
| Narration, pinned language | \nMultilingual 2 | \nlanguageCode | \n
| Fast or bulk speech | \nTurbo 2.5 | \nsame shape, lower cost | \n
| Music from one prompt | \nMusic v2 | \nprompt, durationSeconds, forceInstrumental | \n
| Structured song | \nMusic Advanced v2 | \nrequired sections array | \n
| SFX and loops | \nSound Effects 2 | \ntext, promptInfluence, loop | \n
| Localize, protect names | \nDubbing v2 | \nfile, targetLang, keyterms | \n
| Localize, speaker control | \nDubbing | \nnumSpeakers, dropBackgroundAudio | \n
| Re-voice a recording | \nSpeech to Speech, Voice Changer | \naudio (up to 5 min) plus a voice | \n
| Clean a noisy voice | \nVoice Isolator | \naudio | \n
The speaking members share one voice contract: voiceId takes a cloned ElevenLabs voice and always wins; publicVoice picks a preset (21 at authoring time) and is ignored beside it; neither set means the Adam preset. Voice Changer is Speech to Speech plus delivery dials (stability, similarityBoost, styleExaggeration, useSpeakerBoost). WAV output exists on the TTS members alone; outputFormat elsewhere is MP3 or Opus, absent on Dubbing and Voice Isolator. seed gives repeatability, except on Sound Effects 2, both Dubbing members, and Voice Isolator. Cost rides the content: text on TTS (40,000 characters at authoring time), durationSeconds on Music v2, sections on Advanced, file or audio on the dubbing and voice members, so dry_run long jobs and member comparisons.
Eleven v3 reads inline audio tags in the text, [whispers], [excited], [sighs], to steer delivery moment to moment across 70+ languages, and carries multi-speaker dialogue. Multilingual 2 and Turbo 2.5 do not read tags: direction there lives in stability, styleExaggeration, and speed (0.7 to 1.2), and languageCode (ISO 639-1) pins the language, a field v3 lacks. Turbo trades expressiveness for cost, roughly half the other two per run at authoring time.
Music v2 takes one prompt (mood, genre, instruments, tempo), durationSeconds (3 to 600 at authoring time, cost impact), and forceInstrumental to suppress vocals: asking in prose is unreliable. Music Advanced v2 requires sections, up to 30 ordered segments at authoring time, each with text, its own durationSeconds (3 to 120), positiveStyles and negativeStyles (up to 10 each), and contextAdherence (high binds a segment to its neighbors, low frees it). Section grammar: square brackets label ([Verse], [Chorus]), curly braces direct (), and plain text is sung as lyrics. Advanced has no instrumental flag, so any plain text will be sung. Neither music member takes numOutputs: one run is one composition, and sung lyrics land differently on every draw, so a batch of takes is several parallel runs with distinct seed values, each priced alone on its duration, which costs what a batch field would have and returns separate files with real boundaries. Two [Verse] blocks in one run to get two takes returns one file with no split point.
Both Dubbing members take an audio or video file with a required targetLang, and the output follows the input kind: video in, dubbed video out. Neither re-animates lips. Source auto-detection is spelled differently: sourceLang: \"auto\" on v2, an empty string on the older member. Pick v2 for keyterms (names, brands, and jargon preserved verbatim through translation); pick the older Dubbing for numSpeakers (0 auto-detects, up to 10), dropBackgroundAudio, disableVoiceCloning, and highestResolution video. Their prices differ several-fold, so dry_run both when either fits. Speech to Speech, Voice Changer and Voice Isolator, by contrast, take audio only, whatever the catalog blurb says: for a clip, pull the track with model_scenario-audio-extract, run the member, then lay the result over the clip in model_scenario-compose-video, the clip as a video layer with mute: true (fixed ids: each is Scenario's single deterministic tool for its operation, so discovery would only re-derive it). The same compositor lays a Music v2 bed under a video; layer contract in scenario-video-assembly.
search with target=\"models\", query=\"elevenlabs dubbing\", public=true. Prefer the newest non-deprecated hit, e.g. model_elevenlabs-dubbing-v2 (a live hit at authoring time: re-discover each session).model_schema_get with that id: fields and defaults before anything else.upload_asset the trailer video (see the scenario skill) to get an asset id.model_run with that model_id, dry_run=true, and the exact parameters={\"file\": \"asset_x\", \"targetLang\": \"es\", \"keyterms\": [\"Aetherfall\", \"Kestrel Squad\"]}. The file drives the price: re-estimate per input.model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.asset_display the result (a video, because the input was) and confirm the keyterms survived; asset_download with no format to save it.publicVoice next to voiceId and expecting it to apply: a set voiceId always wins.forceInstrumental; on Music Advanced there is no flag and plain section text is always sung.numOutputs on the music members, or repeating one long run hoping for a different take without changing seed: same seed and settings reproduce the same music.