Audio similarity, as a service.
The matching engine behind everysong.app is available as a service. Labels, music supervisors, stock-music platforms, sync agencies, and indie streaming services can plug into the same pipeline that runs on the consumer site, configured against your own catalogue rather than ours.
The consumer product at everysong.app is a public demo of the engine: 3,382 Creative-Commons tracks, $5 lifetime access for individual creators. The same engine scales to millions of tracks across any catalogue you bring. This page is the partnership surface.
The engine, briefly
Same stack as the consumer site. No proprietary black box.
librosa + pyloudnorm
BPM, musical key, LUFS integrated loudness, spectral centroid, spectral rolloff, stereo width, zero-crossing rate, vocal/instrumental detection. ITU-R BS.1770 compliant for loudness. Open-source, well-validated.
LAION CLAP (laion/clap-htsat-unfused)
512-dimensional audio embeddings via a pre-trained contrastive language-audio model. Captures timbral and stylistic similarity beyond raw signal stats. Cached per-track; one-shot inference at ingest, near-zero cost per match query.
Cosine similarity over L2-normalised matrix
In-memory numpy. Scales to ~100k tracks per process before warranting sqlite-vec or pgvector. Sub-millisecond per match at consumer-site scale; well under 100 ms at 50k-track catalogues.
13 named traits + per-trait deltas on every match
Eight signal-processing traits (BPM, key, LUFS, spectral, stereo width, zero-crossing, vocal/instrumental) labelled GREEN. Five ML-classifier traits (energy, valence, danceability, acousticness, instrumentalness) labelled AMBER. Reliability tier is shown next to each value. No fake numbers.
Full methodology and the underlying paper references are on /how-it-works.
Industry use cases
Brief-to-catalogue matching
Drop the reference track from the director's temp score. Get the closest catalogue matches ranked by acoustic similarity, with full trait readouts so you can defend the choice in a brief response. Works across multi-label catalogues if you have aggregated rights.
Sonic comparables for unsigned artists
Run a new demo against your roster. Find the closest internal sonic neighbours to position a signing decision: who they sound like, who they would tour with, what playlist context they fit. Trait deltas show exactly why the comparison holds.
Search-by-audio for your existing customers
Augment text-and-mood search with audio similarity. Your customers upload their video temp track, your platform surfaces the catalogue tracks that match. White-label or co-branded. Reduces the time a customer spends browsing, increases conversion to license.
Clearance-friendly alternatives, faster
When the brief calls for a specific copyrighted track that won't clear, find sound-alike candidates from your representation roster in seconds, not days. Score them against the original on 13 axes so the client sees objective comparisons.
Recommendation that listens to the audio, not just the metadata
Existing recommendation engines lean heavily on user behaviour and editorial metadata. Audio-content similarity adds a signal that works for cold-start tracks (new releases with zero listener data) and rare-genre catalogues where collaborative filtering is thin.
Bring your own catalogue
The everysong public catalogue is 3,382 Creative-Commons tracks. The engine is catalogue-agnostic. We ingest yours.
Ingest pipeline, per track:
- Audio file (MP3, WAV, FLAC, M4A, OGG) or streaming URL
- librosa extracts ~30 signal features (~5 sec per track on a single CPU)
- CLAP generates the 512-dim embedding (~10 sec per track on a single CPU; ~1 sec on a single GPU)
- Optional: pull your existing metadata (license, artist, ISRC, release date, etc.) and join it to the trait vector
- Indexed for query in your dedicated environment, isolated from the public catalogue
End-to-end for a 50,000-track catalogue on a single GPU: roughly 14 hours, one-shot, never re-run unless you re-encode the audio. Ongoing add-track latency: under 30 seconds per new release.
How you plug in
| Integration | What it is | Best for |
|---|---|---|
| API access | POST audio, GET matches. JSON in, JSON out. Documented at /api-docs. Public CC catalogue available immediately; custom catalogue ingested per partnership. | Internal tools, custom workflows, your own UI |
| Hosted UI (co-branded) | Your domain, your colours, the everysong matching engine underneath. Catalogue isolation guaranteed. | Customer-facing search, sales demo tools |
| White-label | Full ownership, your branding, no everysong attribution. Source available on shutdown (same promise as the consumer product). | Larger platforms with brand requirements |
| On-premise / private cloud | Self-hosted in your infrastructure. We ship the container, you run it. CLAP weights and signal-feature code are open-source. | Labels with strict catalogue confidentiality requirements |
Pricing
Custom per partnership. No public price list because the variables (catalogue size, query volume, integration depth, hosting region, support tier) span a wide range. As an order of magnitude:
- API access for a 50k-track catalogue with ~10k matches/month: starting around $400/month, EU-hosted, shared infrastructure.
- White-label integration with custom catalogue (~50k tracks), ~100k matches/month, co-branded UI: roughly $1,200 to $2,000/month depending on hosting and support.
- On-premise license with full source and a year of support: one-time five figures, negotiated.
Open-book pricing as a principle. Operations are run by one person currently (Zara, Netherlands) so the cost basis is transparent and the markup is modest. Volume discounts on commitment. Partnership structures (revenue share, equity-for-platform) are open for discussion if the alignment is right.
Honest constraints
The boring stuff that matters in due diligence:
- Solo-founder operation. One person on the codebase, on the infrastructure, on the partnership conversations. Partnership terms include explicit handover provisions if continuity is a concern.
- Catalogue isolation. Your catalogue is never co-mingled with the public CC catalogue or any other partner's. Separate database, separate process, separate trait vectors. Confidentiality by architecture.
- Open-source on shutdown. Same brand promise as the consumer site. If everysong stops operating, the matching engine and your integration code go open-source. Your catalogue stays yours; the analysis pipeline survives.
- EU-based and GDPR-shaped. Default hosting region is EU (Fly's Amsterdam or Frankfurt). Data processing agreement available. No third-party transfers without explicit partnership terms.
- No exclusivity contracts. You can build your own matching engine in parallel or move to a competitor at any time. Standard 30-day partnership termination clause.
What we will not do
- Train your catalogue's embeddings into ours. Your audio stays in your environment.
- Sell your catalogue metadata to anyone. Ever.
- Lock the matching engine behind a proprietary format. Output is portable JSON, embeddings are exportable.
- Promise accuracy numbers we haven't measured. Trait reliability is colour-labelled (GREEN/AMBER) as it is on the consumer site.
Partnerships, integrations, RFPs, exploratory chats. Replies within 48 hours during business days (Netherlands timezone, CET/CEST).
[email protected]For consumer-tier ($5 lifetime) support: [email protected]. For refunds: [email protected].
See also
- API documentation: endpoint reference for
/api/v1/match,/api/v1/catalogue/info,/api/v1/me. cURL examples, response shapes, rate-limit tiers, error envelope. - How the engine works: the 5-hop pipeline, the 13 audio traits, GREEN vs AMBER reliability tiers, the catalogue specifics.
- About everysong: who built it, the brand promises, the open-source-on-shutdown commitment.
- The consumer site: $5 indie demo of the same engine.