Sonilo raises $11M to score video automatically, and library composers are the ones in its path

Sonilo, a generative audio start-up in San Francisco, announced an $11 million round on 14 Oct 2026, led by B Capital with Redpoint participating. The pitch: upload a clip and its model writes music and sound effects that follow the edit, without the user describing the track in words. The founders came from TikTok, and the customers it names are short-form creators, short-drama studios and the developers building AI video tools. Digital Music News carried the announcement on 22 Oct.

What does Sonilo actually do?

The release describes a proprietary multimodal system Sonilo calls the "Sound World Model", which links video understanding, audio generation and timing. You can give it footage, a text description, or both, and get back what the company calls a commercially ready, synchronised soundtrack. The same platform does AI dubbing with lip-sync, aimed at short-drama studios localising into other languages, and Sonilo says a consumer app is due in Q4.

The team is the story investors are buying. CEO and co-founder Shawn Song was a multimodal AI tech lead at TikTok; CTO Alex Yin holds a PhD in computer music from the University of York and worked alongside Song at TikTok; COO Keli Li built and led TikTok's global music operation, with a background in music licensing. That is a team that knows both the model side and how short-form platforms clear music.

Distribution is already wide. Early partners named in the release are TapNow and ComfyUI, and the Sonilo homepage lists plugins for Adobe Premiere Pro, Unity, Godot, Epic's Fab, Roblox and Defold. The same page claims, as of September 2026, more than 700 brands using it and 5.3 million "copyright-risky soundtracks replaced". Those are the company's own figures, and none of them has been checked independently.

What did it train on?

This is the question that matters most, and Sonilo answers it more directly than most AI music companies. In a May 2026 post, it said it had licensed Shutterstock's music catalogue to train its v1.0 model, and called the deal Shutterstock's first with a video-to-music AI platform. Its licensing page now says the music models are trained "only on professionally licensed music, including the Shutterstock catalog", and the funding release says licence fees and revenue share flow to rights holders and artists.

Three caveats sit underneath that.

• "Including" isn't a full list. Sonilo doesn't publish its full list of training sources. The funding release says master and publishing deals with "major rights organizations" are still being finalised, "expected to close in the coming weeks". So the current model's licensed catalogue is narrower than the one Sonilo is negotiating for.
• Sound effects are separate. The licensing page says the Shutterstock music deal isn't an inventory of its sound-effects training data; it only states that those sources are "compliant", without naming them.
• Paid "through" whom? Sonilo's post says the artists behind the catalogue consented and are compensated. Its about page puts the money as going to rights holders "and through them to the artists". What a Shutterstock composer actually receives depends on that composer's contract with Shutterstock, and neither company has published the split.

What do buyers get, and what can't they do?

Pricing on the homepage starts with a free tier (non-commercial, preview-quality exports; Sonilo's licensing page asks for a "Music by Sonilo" credit on anything shared). Pro is $11.99 a month billed yearly ($143.88) and adds commercial use for ads, client work, monetised channels, apps and indie games. Premium is $23.99 a month billed yearly and adds film, TV, broadcast and studio games. One condition stands out: according to the licensing page, no plan allows releasing a generated track through a distributor to Spotify, Apple Music or other streaming services. Sonilo is selling background music for video, not records.

It is also careful about what licensed training does and doesn't guarantee. The licensing page says it doesn't guarantee copyright protection or freedom from Content ID claims, while noting that none of 2,000 tracks Sonilo uploaded to YouTube in its own test was claimed.

Who is most exposed?

Not songwriters with hits. The people this replaces are those who write low-budget cues: library and stock composers, producers who score brand content for social video, and the music editors who cut a library track to the length of a 30-second spot. Sonilo's own about page frames the product as an alternative to "digging through endless libraries".

There is a number to put on that worry. A CISAC study published in December 2024 estimated that 24% of music creators' revenues are at risk by 2028. It also projected that generative AI music would account for around 60% of music libraries' revenues by then, far more than the roughly 20% it projected for streaming. Of the two music markets the release singles out, libraries are where it expects AI to take the biggest share, and frame-synced video scoring goes straight at them.

What we think

Sonilo deserves credit for training on licensed music, publishing a licensing page with real limits on it, and ruling out distribution to streaming services, which should keep its output from competing with artists for streaming royalties. That is a better starting point than training on scraped music and settling later.

But licensed training doesn't settle the economics, and the overlap here is unusually direct. As far as Sonilo has disclosed, its main named music source is a stock-music catalogue. That is music written to be cheap, cleared once, and laid under video. A model trained on it, sold from $11.99 a month and built into Premiere and game engines, competes in exactly the market that catalogue's composers earn from. In effect, those composers are being paid a share of a licence fee to train the product that will undercut their next library placement. Whether that is a fair trade depends entirely on a figure nobody has published: what each contributor earns from the training deal compared with what they lose in placements. (This is our analysis, not a claim either company makes.)

Second-order effects to watch:

• Winners: short-form creators and small studios, who get timed music without licence worries; Shutterstock, which earns from the same catalogue twice; and short-drama studios that want localisation at scale.
• Losers: stock and library composers, and freelance editors whose job was fitting music to picture. The first jobs to go are the fiddly timing work, not the creative brief.
• The leverage question: the major-label and publisher deals Sonilo says are weeks away would be the first sign of how labels and publishers price video-scoring models compared with song generators. If they include per-output payments rather than a flat training fee, that would be a model library composers should demand too.

If you write library music, now is the time to read your library agreement for any AI or data-licensing clause, ask your library whether your tracks are in a training deal and what you are paid for it, and opt out where the contract allows.