What the 13 audio traits actually tell you about a song.

Every everysong readout starts with a card of 13 numbers describing the audio you uploaded. Most people glance at the BPM and key and skip the rest. That's a missed opportunity: the other 11 traits each answer a useful question about the song, from "is this mix actually loud enough?" to "is this track going to feel like dance music or chill electronic on the dance floor?". This is the no-jargon guide to what each trait measures, what counts as a typical value, and what to do with it.

Two ground rules before we start. First: eight of the thirteen traits are "GREEN" tier: they're measurements made by signal-processing algorithms with decades of academic research behind them. You can quote them in a mix note or a brief without hedging. Second: five traits are "AMBER" tier: they're outputs of pre-trained ML classifiers. Trust them as a useful second opinion, not as ground truth. The label is shown next to each reading in the readout so you don't have to remember which is which.

The 8 GREEN traits: signal-processing measurements

The 5 AMBER traits: ML-classifier outputs

These are the "mood-ish" descriptors. The numbers are useful directional indicators but should be treated as second opinions, not measured facts. They come from pre-trained machine-learning models that have been trained on millions of labelled tracks: they're statistical guesses, not measurements.

How to actually read a readout

When you upload a song to everysong, you'll see all 13 of these numbers across the top of the card. Here's a quick "is this reading sane?" check:

1. Glance at the BPM first. If it looks wildly off (e.g. it says 200 BPM but you can clearly tap a slow 100 BPM beat), the engine probably did half/double misdetection. The matches will still be useful but biased toward the misdetected tempo.
2. Check the key. If you have absolute pitch or know the song's key, verify. Most pop song keys are easily checkable against the chord patterns on Ultimate Guitar or similar.
3. Look at LUFS. If you're producing a track yourself and trying to match the reference's loudness target, this is the number you'll consult most often.
4. Use spectral centroid + rolloff together. Two tracks with similar values in both are almost certainly going to feel similar in their high-frequency character.
5. Treat AMBER traits as a tie-breaker. When you have two candidate matches with similar GREEN profiles, the AMBER traits (energy, valence, danceability) help you pick the one with the closer mood.

Worked example: matching an acoustic indie folk track

Pretend you uploaded a quiet acoustic indie folk song: solo voice, fingerpicked guitar, no drums. A reasonable readout might look like:

BPM:              82
Key:              D minor
LUFS:             -18.5
Spectral centroid: 1,820 Hz
Spectral rolloff:  4,400 Hz
Stereo width:     0.55
Zero-crossing:    1,100 / sec
Vocal/inst:       vocal
Energy:           0.32
Danceability:     0.41
Valence:          0.28
Acousticness:     0.87
Instrumentalness: 0.05

Reading this: slow tempo, minor key (so probably melancholic), moderately quiet (-18.5 LUFS is below platform target: will get normalised UP slightly), darker spectral profile (centroid below 2 kHz), reasonably stereo, low percussive content, sung vocal, low energy, low danceability, sad-leaning emotion, strongly acoustic, definitely vocal-led.

The matches you'd see in this case would skew toward: slow tempo acoustic singer-songwriter tracks in minor keys. Other indie folk, some ballads, possibly some acoustic blues or chamber-pop. The engine would NOT return upbeat dance tracks because the energy, danceability, and acousticness all push hard against that.

Worked example: matching a deep house track

Pretend instead you uploaded a 4/4 deep house track:

BPM:              124
Key:              G minor
LUFS:             -8.2
Spectral centroid: 2,950 Hz
Spectral rolloff:  9,200 Hz
Stereo width:     0.32
Zero-crossing:    2,400 / sec
Vocal/inst:       instrumental
Energy:           0.78
Danceability:     0.88
Valence:          0.55
Acousticness:     0.05
Instrumentalness: 0.92

Reading this: standard house BPM, minor key (deep house is mostly minor), loud master (well above platform target), bright spectral profile, wide stereo, percussive (kick + hat content), instrumental, high energy, very danceable, emotionally neutral-to-positive, fully synthetic.

The matches would skew toward: house, techno, deep electronic, possibly some downtempo electronic for the wider net. Wouldn't return ambient (too low energy mismatch) or acoustic folk (acousticness mismatch). The danceability and instrumentalness traits both push in the same direction here.

Why these 13 specifically

v1 deliberately ships only these 13 because each one is well-validated by years of academic music information retrieval research (the GREEN tier) or trained on enough labelled data to be reliably useful (the AMBER tier). The landing page mentions a wider "atlas" of ~190 experimental traits: those are aspirational, marketing-honesty about what the engine could theoretically compute, not features that v1 actually ships. We deliberately don't show numbers we haven't validated.

One trait we considered and dropped: genre classification. Pre-trained genre classifiers leak Free Music Archive's training distribution heavily, so a track that's clearly "deep house" to a human ear can get classified as "electronic > experimental" because that's how FMA's labelling skews. Until we can run the model against a cleaner, broader genre dataset, the genre column stays out of v1.

What to do with all this

Three practical uses for the 13-trait readout:

1. As a mix-reference comparison. Upload your work-in-progress next to a reference you're chasing. The deltas tell you where you've drifted: LUFS, spectral balance, stereo width.
2. As a match-quality sanity check. Before trusting the 20 CC matches the engine returns, verify the 13 traits "look right" for your reference. If the BPM is half-time misdetected, the matches will be off.
3. As a song-tagging system for your own library. If you keep a personal catalogue of demos, run each one through everysong and the 13-trait readout becomes a quantitative tag set. Filter by "all my demos with valence < 0.4 and danceability > 0.7" and you've got your sad-club-bangers shortlist.