An audio engine with a music writer's ears. Tune her to a song.
Measured on signalVera is my own audio engine. I built her because I wanted to work out how and why a track hits the way it does. She runs on my laptop on free everything, and she labels every single thing she tells you.
Tempo, key, loudness, the energy curve. Straight signal processing, so trust these.
Genre from sound, the feel, danceability, the notes. Model and rule based: check them against your ears.
Year, album, credits, catalogue BPM and genre tags, looked up in Deezer and MusicBrainz.
Tempo is marked against the catalogue BPM (510 of the songs have one). Genre is checked on songs her genre model never saw during training. Published as measured, misses and all. When she recognises a song, tempo, year and genre come from the catalogue instead, which is why searching is the best way in.
She was giving every song the same tags, because the model behind her has favourite words it reaches for whatever it hears. Billie Jean and Adele both came back as blues and "good for meditation". Now each tag is measured against how typical songs score (about 230 chart previews across 18 genres), so she only says what's distinctive about this one.
She also used to guess the decade from the sound and got it right 1 time in 5 (Bohemian Rhapsody went in the 1950s). Now the year only comes from the actual release. And she stopped calling pop songs 5/4.
Then she sat a blind exam of 1,137 songs. It caught her calling 27% of everything "latin" and 21% "blues", so genre is now learned from about 900 labelled songs instead of guessed from word prompts: right family 68%, and no single genre takes over. A new tempo picker stopped her hearing 120 BPM as 81. And she no longer calls a song "programmed", because her timing measure couldn't tell a drum machine from a tight band.
Updated 17 November 2026. Same test each time, old version of her against new, so you can see what actually moved.
The exam above is the big one: 1,137 songs she'd never heard. The 30-song set is small, so treat it as a trend, not a trophy.
Search results come straight from Deezer's public catalogue with a 30 second preview, which she listens to. Uploads are fingerprinted and matched against AcoustID and MusicBrainz. Credits and community genre tags come from MusicBrainz where its editors have filled them in, so a missing credit means undocumented, not that nobody did the job. Listen links go to Apple Music and Deezer directly and to a search on Spotify and YouTube Music. No Spotify anywhere in her analysis.
She detects the chord progression and matches it against a curated list of historically significant progressions, naming the family and where it comes from. She runs it twice, on the whole mix and on the bass alone, to catch a twelve bar blues under a wall of guitars. It only shows when she finds a confident match.
She's the engine inside everysong, where she finds free music that sounds like the song your edit was cut to, maps the beat for the whole track and checks whether a swap will keep your cuts on time.
Every readout is the same versioned JSON shape (vera.version), signed so you can prove it came from her. Built on permissively licensed open source, no GPL. Her algorithms, prompt sets and taxonomies are mine.
Running your own copy of the engine? Point this page at it: