> 06_HOW_IT_WORKS.MAN

How everysong works.

You drop an audio file. We return 13 audio traits and 20 ranked Creative-Commons matches. The whole pipeline runs in 10–15 seconds (30s on the very first call while the model wakes up). This page is the unhurried version — what each hop actually does, which traits are trustworthy, and where the limits are.

If you got here looking for a way to find royalty-free music that sounds like a specific song, you're in the right place. People also describe everysong as a music similarity search, an audio fingerprint tool, a copyright-safe music finder, a creative-commons alternative finder, a free Shazam-style tool that returns similar tracks instead of identifying your upload, and a BPM / key / LUFS analyzer with a built-in match engine. They're all describing the same loop: upload → 13-trait audio fingerprint → 20 closest CC-licensed neighbours.

The five hops

01 · INGEST
Drop a file

MP3, WAV, FLAC, M4A, or OGG up to 30 MB. The file lands on the API just long enough to analyse, then it's deleted. We never keep the audio.

02 · ENCODE
Extract signal features & embedding

librosa pulls low-level signal features (BPM, key, RMS energy, spectral centroid, spectral rolloff, zero-crossing rate, stereo width). In parallel, the CLAP audio encoder produces a 512-dimensional embedding — a learned representation that captures vibe-level similarity beyond the raw signal stats.

03 · DECODE
Map embedding → 13 interpretable traits

The 512-dim embedding is too abstract to read, so we project it onto 13 named axes you can actually reason about: tempo, key, loudness, energy, valence (happy/sad), danceability, acousticness, instrumentalness, liveness, speechiness, brightness, complexity, mode. Stored as JSON alongside the file's signal stats.

04 · MATCH
Cosine similarity over 3,382 CC tracks

Your embedding gets compared against the catalogue of 3,382 Creative-Commons licensed tracks — sourced from Free Music Archive, ccMixter, and Jamendo. The 20 closest by cosine distance are returned, ranked, with the per-trait deltas so you can see why they're close.

05 · ACT
Browse, stream, share

You land on a readout page: 13 trait values up top, 20 ranked matches below, license badge on every match (CC0, CC BY, CC BY-SA — never BY-NC, never restricted), preview button, and a deep link to the source page on FMA / ccMixter / Jamendo. There's a permalink so you can share the readout. That's the whole loop. No upsell.

The 13 traits we ship

v1 ships 13 trait readouts, colour-labelled by how much you should trust them. The boring stuff (signal processing) is solid. The mood-ish stuff is a useful second opinion.

01 BPM / tempo GREEN
02 Musical key GREEN
03 Loudness (LUFS) GREEN
04 Spectral centroid GREEN
05 Spectral rolloff GREEN
06 Stereo width GREEN
07 Zero-crossing rate GREEN
08 Vocal / instrumental GREEN
09 Energy AMBER
10 Danceability AMBER
11 Valence (happy/sad) AMBER
12 Acousticness AMBER
13 Instrumentalness AMBER

What "GREEN" and "AMBER" actually mean

GREEN

Signal-processing traits. BPM, key, LUFS, spectral stats. Decades of published research underneath. Well-validated on standard MIR benchmarks. These are the numbers you can quote in a mix note or a brief without hedging.

AMBER

Mid-level descriptors. Mood, energy, danceability, acousticness. Pre-trained classifiers make educated calls based on millions of labelled examples — but they're statistical, not ground truth. Treat as a useful second opinion, not as gospel. The label is right next to the reading so you never have to guess.

RED

No red-tier traits ship in v1. We considered genre-classification and "instrument detection beyond vocal/instrumental" — both either leak training-set bias or aren't reliable enough to publish a number for. They'll show up when they're real.

Who actually uses it

For crate-diggers

Find the song that sounds like the one stuck in your head.

Point at a track. Get 20 sonic neighbours you can actually hear the resemblance in. No genre tags, no "people also listened to."

For creators

Score your video, podcast, or stream — legally, for free.

YouTubers, podcasters, streamers, indie game devs: upload a reference you love, get CC-licensed matches with the license shown up front. Always verify the license at the source before publishing — Creative Commons is freely usable, but not strike-proof if you ignore the attribution rules.

For producers

A/B your mix against the reference, honestly.

Side-by-side trait readout: where you match, where you drift. BPM, key, LUFS, spectral balance, stereo width — all the boring stuff that matters at mastering time.

For the curious

Just geek out over a song you love.

Why does this song hit? Break it open. See the key change, the tempo drift, the moment the reverb opens. A magnifying glass for music.

Honest answers to the obvious questions

> Why $5?
It covers the compute per user (~$1.20/mo amortised), and the rest is coffee. If I charged $20/mo I'd feel like every other tool. One time, paid once, lifetime access — that's the brand promise.
> How is this sustainable?
Flat fees plus a side commercial API for studios that need bulk access. 20,000 × $5 funds a fine solo project and a lot of coffee.
> Can I use it commercially?
Personal and freelance work, yes — that's exactly what it's for. High-volume or embedded-in-product use, talk to me at [email protected].
> What if the model gets better?
You get the new version. Same five bucks, no upgrade charge. Lifetime means the model family, not a frozen v1.
> What if you shut down?
The analysis code goes open-source. Your local CSV exports keep working. That promise is in the terms.
> Where are the published benchmarks?
I won't publish benchmark numbers I haven't actually run. After ~1,000 real public readouts I'll post an eval report — against MIREX-style test sets for rhythm/key, and held-out listener agreement for the AMBER traits. Until then: trust the colour, check the reading.

How everysong compares to other audio tools

People often ask how everysong sits next to other audio tools. The short answer: it's purpose-built for indie creators who need a copyright-safe replacement for a copyrighted reference track, at $5 once instead of a subscription. Here's the long answer:

vs. Shazam

Shazam identifies the recording. everysong finds tracks that sound similar.

If you upload "Bohemian Rhapsody" to Shazam, it tells you it's "Bohemian Rhapsody". If you upload it to everysong, you get 20 Creative-Commons tracks that share its BPM, key, energy and overall feel — songs you can actually use in your video, podcast or game without paying licensing fees. Different tool, different problem.

vs. Cyanite, AIMS, Musiio

Industry audio-AI platforms aimed at labels and supervisors. Subscription, hundreds-to-thousands per month.

These are excellent products and substantially more comprehensive than everysong — they handle catalogues of millions of tracks for music supervision, sync licensing, and label A&R. But they're not for the indie YouTuber who just needs a free track that sounds like the demo they put on their rough cut. everysong is the $5 indie-tier version: smaller catalogue, narrower features, no subscription.

vs. Mubert, Soundraw, Suno, Udio

AI music generators. They invent new tracks. everysong matches existing ones.

Generators are great when you want exactly-fitting bespoke audio. The trade-offs: AI-generated music has unsettled copyright status, often sounds generic, and rarely matches the specific feel of a reference. everysong returns real recordings by real humans, with clean Creative-Commons licenses and verifiable provenance.

vs. Epidemic Sound, Artlist, Soundstripe

Royalty-free subscription libraries. Big catalogue, monthly fee.

Subscription libraries are the right choice if you publish a lot and want a constant tap. everysong is for the case where you have this specific reference track and want a free, license-safe match — not a curated mood-based playlist.

vs. YouTube Audio Library

Free, curated, but no similarity search.

YouTube's library is great for browsing by mood and genre. everysong starts from your reference track and goes outward by audio similarity. Different access pattern.

What people search for to find this

People stumble onto everysong via dozens of different phrasings. If any of these sound like you, you're in the right place:

What's in the Creative-Commons catalogue

v1 ships with 3,382 carefully-curated CC tracks pulled from three major free-music sources:

License coverage: CC0 (public-domain dedication, no attribution needed), CC BY (attribution required, commercial use allowed), CC BY-SA (attribution + share-alike). We exclude BY-NC tracks because they prohibit commercial use — most creators end up needing commercial-safe music. There's also a "strictest only" filter that hides BY-SA tracks if you want zero share-alike obligations on a client deliverable.

Other names for this kind of tool

For posterity and Google: this category gets called music information retrieval, audio similarity search, music recommendation by audio, audio fingerprinting and matching, content-based music retrieval, audio embedding similarity, CLAP-based music search, music feature extraction, audio MIR (Music Information Retrieval), and informally music DNA or audio genome. Academic papers tend to say "MIR"; vendors tend to say "audio AI"; creators tend to say "find similar songs".

Ready to find free music that sounds like your song?
$5 once · lifetime access · 30-day no-questions refund
▶ Pay $5 · Unlock Everything
← everysong