Summary: ElevenLabs documents a POST /v1/music/video-to-music endpoint that generates background music directly from one or more video files, with optional style tags, SFX extraction, and C2PA provenance signing. For builders stitching AI video pipelines, it removes the manual score-and-mix step. Here’s what the API actually offers and where the constraints sit.
What the endpoint does
You upload one or more videos via a multipart form, and the API returns a generated audio file that matches the footage. The documentation specifies a hard ceiling: up to 10 videos, 200MB combined, 600 seconds total. Videos are combined in order, so the track’s structure follows your edit sequence.
Two optional steering inputs shape the result: a free-text description (up to 4100 characters) and a list of up to 10 style tags like upbeat or cinematic. Simple, but enough to build a UI on top without inventing your own taxonomy.
The flags that matter for production pipelines
Three booleans stand out:
add_sfxpulls SFX cues from the video into the generation.separate_tracksreturns a ZIP with music and SFX as separate audio files instead of one mix — the right default if your editor wants to control levels downstream.use_audio_as_referenceanalyzes the source video’s audio for genres, BPM, instrumentation, mood, and arc, then uses those as generation characteristics. That’s a useful path when you’re replacing licensed temp music with something tonally identical.
There’s also sign_with_c2pa, which signs the generated mp3 with C2PA content credentials. I covered why that matters in the Spotify verified badge piece: as AI music floods distribution channels, provenance metadata is becoming table stakes rather than a nice-to-have.
Formats and access tiers
The output_format parameter covers MP3 at several bitrates, PCM from 8kHz to 48kHz, μ-law and A-law, and Opus. Two tiers gate the good stuff: 192kbps MP3 needs Creator or above, and 44.1kHz PCM needs Pro or above. If you’re feeding audio into a Twilio voice flow, the μ-law option is explicitly called out for that use case.
Regional endpoints exist for US, EU, India, and Singapore residency — worth planning for early if you have data locality requirements, since changing base URLs later touches auth and logging configurations.
Where it fits in a video generation stack
This endpoint assumes you already have footage. If you’re generating that footage with AI models too, the pieces now slot together: pick a model, generate clips, concatenate, then send the result here for a score. I’ve written before about how Seedance 2.5 pushes clips to 30-second one-takes — endpoints like this are what make those clips shippable without a composer or a stock-music subscription.
One practical note from the docs: combining multiple videos server-side can significantly increase request duration, and ElevenLabs recommends pre-concatenating your videos when possible. If your pipeline already edits footage before scoring, keep that order and send a single file.
Grounded next step
The fastest way to evaluate this is a thin script: one clip, description set to your brand’s audio direction, separate_tracks on, and C2PA signing enabled so you can inspect the metadata. A 403 means your subscription tier doesn’t cover it; a 422 usually means the video payload failed validation. The supplied documentation page doesn’t specify pricing beyond tier gating, so check the subscription terms before committing to a per-video cost model.
Sources
View the original post on X