Elevenlabs Music

ElevenLabs Music production skill

Use this skill to turn a content brief into a usable music asset with ElevenLabs Eleven Music. Treat it as a production workflow: choose the right generation route, write legally safe musical direction, preserve artifact custody, and review the result against the edit before publishing.

Source-backed facts in this skill were verified on 2026-07-10 from official ElevenLabs documentation and terms. Re-check live docs before relying on pricing, limits, model defaults, licensing terms, or endpoint schemas in a paid or public production.

Hard boundaries before prompting

Do not submit prompts, lyrics, references, uploads, or metadata that name or clearly target a protected work or identifiable artist. ElevenLabs Music Terms prohibit Music inputs that include any artist real/stage name, songwriter real/stage name, song title, album title, music publisher name, label name, or substantial/distinct song lyric excerpt intended to reference a song. They also prohibit prompting likely infringement and misleading mimicry of an identifiable recording artist.

Use genre, era, region, instrumentation, production language, tempo, mood, arrangement, mix references, and functional intent instead:

• Unsafe: "Make it sound like [artist]'s [song title]."
• Safer: "late-1970s glossy disco-pop, four-on-the-floor drums, octave bassline, bright string stabs, celebratory chorus, 118 BPM, no artist imitation."
• Unsafe: "Write a hook like this copyrighted lyric line..."
• Safer: "uplifting two-line chorus about starting fresh, original lyrics only."

For commercial release, check the active plan and Music Model-Specific Terms. As verified 2026-07-10, self-serve plans differ by generation/download limits, attribution, streaming rights, API features, concurrency, and permitted media. Self-serve media rights were documented as allowing online/offline commercial use except film, TV, radio, and Studio Games; Enterprise Music was documented as allowing all online/offline commercial use. Do not imply a client has broadcast, film, TV, Studio Games, reseller, or music-library/repository rights unless their plan/contract explicitly grants them.

Capability snapshot

Documented Eleven Music capabilities include text-to-music generation, optional vocals or instrumental output, multilingual lyrics, section/lyric editing, composition plans, inpainting with music_v2, video-to-music, upload for inpainting, and stem separation. API access is for paid users.

Key API facts verified 2026-07-10:

• Basic compose endpoint: POST /v1/music.
• Detailed compose endpoint: POST /v1/music/detailed, returning audio plus composition plan and song metadata.
• Composition-plan endpoint: creates a plan without generation cost, though rate limits apply.
• Video-to-music endpoint: accepts up to 10 videos, total combined video size up to 200 MB, and total video duration up to 600 seconds.
• Prompt and composition_plan are mutually exclusive in compose calls.
• music_length_ms with prompt generation accepts 3,000-600,000 ms.
• Official duration documentation was not perfectly aligned: the API reference documented 3,000-600,000 ms for music_length_ms, while the overview and pricing pages also described a 5-minute generation/duration limit. Verify live account/API behavior before promising a cue longer than 5 minutes.
• prompt accepts up to 4,100 characters.
• model_id values documented for music include music_v1 and music_v2.
• On the overview page, music_v2 was documented as the UI default while API default remained music_v1 during a transition period. Set model_id explicitly instead of depending on defaults.
• Default output_format="auto" selects MP3 format according to the model: documented as mp3_44100_128 for v1 and mp3_48000_192 for v2.
• seed can improve consistency with the same parameters, but exact reproducibility is not guaranteed and outputs may change across system updates.
• force_instrumental=true guarantees an instrumental result with prompt generation and cannot be used with composition plans.
• API pricing was documented as $0.15/minute for Eleven Music on the API pricing page. Confirm before estimating cost.

Choose the generation route

For fast creative exploration, use a prompt with explicit duration and model_id="music_v2" unless a project requires v1 compatibility. Ask for 2-4 variants when using the web UI; with API, create separate generations and track seeds, prompt versions, and selected take.

For an edit-locked cue, use a composition plan. Plans give explicit chunk durations, section labels, lyrics placement, positive styles, negative styles, and context adherence. Use plans when the cue must hit voiceover gaps, chapter changes, trailer beats, looping regions, or a precise chorus/outro.

For scoring an existing cut, use video-to-music when the footage itself should guide the background music. Still provide a concise description and up to 10 style tags so the model understands brand tone and edit purpose. Combine videos beforehand when possible to reduce request duration and avoid ordering mistakes.

For revision after a good take, use music_v2 inpainting instead of regenerating the entire song. Generate with store_for_inpainting=true or upload an owned/non-infringing file, keep good slices as audio-reference chunks, and regenerate only weak sections.

For mix control, request or derive stems when you need to duck vocals under narration, remove drums during dialogue, create a social cutdown, or hand off to a sound mixer. Expect high latency on stem separation for longer files.

Prompt construction

A strong Eleven Music prompt answers five questions in musical language:

1. What is the asset for? State the production function: "15-second pre-roll ad bed," "loopable game menu music," "60-second explainer background under narration," "end-card sting."
2. What should the audience feel? Use emotional and tonal words, not just genre names.
3. What should be heard? Specify instrumentation, rhythm section, tempo/BPM range, key or harmonic color when useful, vocal presence, and mix density.
4. How should it move over time? Include structure, transitions, rise/fall, drops, breaks, stingers, or section timing.
5. What must not happen? Add negative direction such as no vocals, no busy lead melody, no copyrighted references, no harsh cymbals, no sudden ending, no artist imitation.

ElevenLabs' best-practices docs state that high-level intent prompts can work, but production prompts should still include constraints that protect the edit: duration, vocal policy, density under dialogue, timing cues, and deliverable context.

Use "instrumental only" in the prompt and force_instrumental=true for prompt-based API calls when vocals would conflict with narration. If lyrics are acceptable, provide original lyrics and timing cues such as "lyrics begin at 15 seconds" or "instrumental only after 1:45." When lyrics are not provided, the model may generate structured lyrics that match the prompt and duration.

For isolated elements, use targeted phrasing: "solo piano," "solo nylon-string guitar," or "a cappella female vocal texture." For a full mix, specify that those elements should sit inside a full arrangement.

Composition plan guidance

Use music_v2 chunk-style plans for most production work. A plan can contain up to 30 chunks. Generation chunks include:

• text: section label in square brackets, original lyrics, or inline directions.
• duration_ms: 3,000-120,000 ms for each generation chunk.
• positive_styles: up to 50 style/direction terms.
• negative_styles: up to 50 terms to avoid.
• context_adherence: high, medium, or low.
• Optional conditioning_ref and condition_strength for reference-conditioned regeneration.

The first chunk strongly sets the overall tone. Give early chunks at least 6-7 concrete style terms until the direction is established. Use context_adherence="high" for seamless edits and medium or low only when a section should noticeably change.

Audio-reference chunks preserve slices of a stored song unchanged:

{
  "song_id": "stored_song_id",
  "range": { "start_ms": 0, "end_ms": 30000 }
}

Conditioning references can guide a generation chunk; the documented maximum conditioning reference length is 30 seconds. Use condition_strength low, medium, high, or xhigh depending on how tightly the new chunk should follow the stored source.