What Actually Makes Music Feel: I Tested My Own Emotion Model on 320,323 Tracks

The Sonic Analyzer scores every track it hears for emotional character. Power, tenderness, nostalgia, wonder, transcendence, and so on. Those scores come from a model I built by hand, mapping audio features onto feelings, weight by weight, over twenty years of deciding what things sound like for a living.

It uses six families of measurement: the seven frequency bands, tonal balance, rhythm, dynamics, compression, and spectral character. About thirty features, each nudging one emotion or another up and down.

I had never tested any of it against anything but my own ears.

The Problem With Marking Your Own Homework

The obvious study is to correlate my emotion scores against the audio features and publish the result. That study is worthless, and it is worth saying why.

My model computes emotion from those features. There is a line in it that says air content raises the transcendence score. If I then measure whether transcendent tracks have more air, I have discovered nothing. I have read my own assumption back to myself with 300,000 decimal places of false confidence. A lot of published audio analysis has exactly this problem buried inside it.

So the model could not be its own judge.

Method

The tracks carry editorial mood tags assigned by a third party with no access to our measurements. 235 distinct tags: melancholic, hypnotic, euphoric, spacey, bitter, anthemic, and a long tail.

I mapped those onto the eleven categories the analyzer uses. A track only counts toward an emotion if every one of its tags falls in that single category. A record tagged both joyful and melancholic is discarded rather than counted twice. That leaves 30,912 unambiguous tracks out of a corpus of 320,323.

For each emotion I compared its tracks against every other track in the set, on each feature, using a Mann-Whitney U test. Effect sizes are rank-biserial correlations, which run from −1 to +1 and do not care about outliers or non-normal distributions. Across all 121 comparisons I applied a Benjamini-Hochberg correction, so the numbers below survive being one of many tests rather than being cherry-picked.

98 of the 121 comparisons were significant at a false discovery rate below 5%. That sounds impressive until you look at the effect sizes, which are mostly modest. With samples this size, almost anything real becomes detectable, so the question worth asking is not what is significant but what is large.

Finding One: It Is Mostly One Axis, And It Is Arousal

Emotion	Tracks	Air >8k	Mid 500Hz-2k	Brightness	Energy	Crest	Compression
Power	1,987	16.94%	23.77%	3,185 Hz	0.327	9.91	0.619
Tension	1,100	15.20%	25.17%	3,122 Hz	0.312	10.26	0.605
Transcendence	1,090	16.18%	22.64%	2,813 Hz	0.308	10.30	0.604
Joyfulness	2,762	16.86%	25.39%	3,111 Hz	0.306	10.48	0.597
Brooding	869	15.58%	25.25%	3,039 Hz	0.301	10.56	0.594
Wonder	1,835	13.31%	25.07%	2,723 Hz	0.293	10.67	0.590
Tenderness	5,244	16.08%	26.13%	2,972 Hz	0.291	10.91	0.580
Nostalgia	969	14.18%	27.44%	2,790 Hz	0.291	10.81	0.584
Sadness	11,418	13.88%	27.19%	2,801 Hz	0.277	11.14	0.572
Peacefulness	3,519	13.16%	25.53%	2,392 Hz	0.268	11.26	0.567

Read down energy, crest factor and compression. They move in lockstep and in opposite directions: as energy rises, crest factor falls and compression rises. Power sits at one end of all three, peacefulness at the other, and everything else lines up between them.

The single strongest relationship in the entire study is peaceful music being darker, at a rank-biserial correlation of −0.311 and a z of −30.1. Nothing else comes close. If a record is meant to feel calm and it measures bright, something is fighting the intention.

Power is the most distinctive emotion overall: energy +0.285, compression +0.285, crest factor −0.285, brightness +0.247, loudness +0.239, all comfortably significant.

Finding Two: Tempo Is the Weakest Thing in the Model

Every feature I tested separates between eight and eleven of the eleven emotions. Except one.

Feature	Emotions separated	Largest effect
Air above 8k	11 / 11	0.237
Bass, brightness, danceability	10 / 11	0.284 to 0.311
Midrange, energy, crest, compression, dynamic range	9 / 11	0.179 to 0.285
Loudness	8 / 11	0.239
Tempo (BPM)	4 / 11	0.100

Tempo separates four, with a largest effect roughly a third the size of everything else. And the four are all at the calm end: peacefulness, sadness, nostalgia and tenderness sit fractionally slower. For power, tension, joyfulness, wonder and transcendence, tempo predicted nothing at all, every one of them failing significance.

Producers talk about tempo as if it sets the emotional temperature of a record. Slow it down to make it sad. At the scale of a commercial catalogue that relationship barely exists, and where it exists it is small. It is also the second time our data has said so: the tonal balance study found tempo did not predict brightness either, at a rank correlation of +0.07.

So Why Does Everyone Believe It?

Because for most of recorded history it was true, and the reason is mechanical.

Pull a record down from 33 and it gets darker. Not metaphorically, measurably: less rotation under the needle, everything shifts down, the whole spectrum drops with it. Push it back up and it brightens again. Anyone who has ever had their hand on a turntable knows this in their fingers. The same was true of tape.

But that is pitch, not tempo. Speed and pitch were welded together, so slowing a record down genuinely made it darker, and darker genuinely reads as sadder. Three generations of engineers learned that association from machines where it was physically real.

Then software separated the two. You can now halve a tempo and leave the spectrum exactly where it was. The mechanism that made the old belief true is gone, and the belief stayed behind.

That is what these numbers are catching. Whatever makes a record feel a certain way, it is mostly not the speed it is played at. It is how hard it is pushed.

Finding Three: Ethereal Records Are Scooped, Not Bright

This is the result I did not see coming, and it is the most useful thing in the study.

I assumed transcendence lived in the air band. Fifteen years of engineers reaching for a high shelf to make something feel vast. What the data says instead:

Transcendent records have the least midrange of any emotion measured. A median 22.64% of spectral energy in the 500Hz to 2kHz band, against 27.4% for nostalgia and 27.2% for sadness. The effect size is −0.243 with a z of −13.7, the second largest in the whole study. Alongside it, bass runs +0.175 above the field.

So the ethereal signature is a scoop: less middle, more bottom, and only modestly more top. Not a brightness move at all. Transcendence was not even significantly brighter than average.

That reframes a common instinct. When something needs to feel vast, the reflex is to add air. The records that actually read that way to listeners got there by taking the middle out. Air is part of it, but it is the smaller part.

Here is how I think about why that works, and it is the only model I have that fits the numbers.

When you listen to a record, brightness is the front wall and the sub is the back wall. Those two extremes are what tell you how big the space is. Everything in the middle is furniture standing between them.

Take the furniture away and the two walls are all you have left, so the room reads as enormous. Push energy to both ends while emptying the centre and you have manufactured distance: far from the front, far from the back, nothing in between to give you a sense of scale. That is the heavens, or the idea of them.

Which is exactly the shape the measurements describe. Transcendent records are not simply brighter or simply heavier. They are hollowed out in the middle with both walls still standing, and the emptiness is what makes the space.

Where I Was Right, and Where I Was Wrong

Right: air does carry transcendence. My model weights air positively toward transcendence, and it holds: effect +0.098, z = +5.5, comfortably significant. It is about half the size of the air effect on power or joyfulness, so the weight is probably too strong, but the direction was correct.

Right: midrange carries nostalgia and sadness. Nostalgia +0.103 and sadness +0.143 on midrange share, both significant. Nostalgia has the highest median midrange of any emotion at 27.44%. If a record needs to feel like memory, the answer is in the middle.

Right: brightness carries joy and power. +0.190 and +0.247.

Probably wrong: air does not carry wonder. I had this weighted positive. It measures −0.086 with a z of −6.2. Wonder is one of the least airy emotions in the set, at 13.31%, which means my model would score a record down on exactly the quality it was labelled for.

I am hedging on this one, because the wonder bucket is roughly two thirds the single tag psychedelic, and psychedelic is closer to a genre than a feeling. So this may be telling me that psychedelic rock is dark, which it is, rather than that wonder is dark. Until that category is re-rated by somebody who is not me, treat it as a flag on the weight rather than a verdict on it.

Wrong: bass does not carry power. Weighted positive, measured negative. Powerful records are bright and lean, not heavy.

Unproven: several features are earning their weight only barely. Dynamic range produces effects around 0.179 at best. Tempo, as above, mostly nothing.

The Tells, Genre by Genre

Genre is the obvious objection: peaceful records skew acoustic, powerful ones skew electronic. It survives the control. Comparing rock only against rock, peaceful rock medians at 2,596 Hz and powerful rock at 3,169 Hz, a 573 Hz gap with nearly five points of high-mid energy between them.

Run it across every lane and something else appears. Genres differ enormously in how much room they give emotion to move.

Genre	Brightest emotion	Darkest emotion	Spread
Instrumental / score	Sadness, 2,198 Hz	Peacefulness, 929 Hz	1,269 Hz
Electronic	Power, 3,639 Hz	Peacefulness, 2,492 Hz	1,147 Hz
Ambient	Sadness, 1,590 Hz	Peacefulness, 720 Hz	870 Hz
House	Tenderness, 3,347 Hz	Peacefulness, 2,689 Hz	658 Hz
R&B / Soul	Joyfulness, 3,027 Hz	Peacefulness, 2,377 Hz	650 Hz
Rock	Power, 3,169 Hz	Peacefulness, 2,596 Hz	573 Hz
Pop	Joyfulness, 3,121 Hz	Wonder, 2,580 Hz	541 Hz
Dance / EDM	Joyfulness, 3,458 Hz	Peacefulness, 2,933 Hz	525 Hz
Country / Folk	Tenderness, 2,988 Hz	Peacefulness, 2,481 Hz	507 Hz
Indie	Tenderness, 2,712 Hz	Peacefulness, 2,226 Hz	486 Hz
Alternative	Tenderness, 2,846 Hz	Peacefulness, 2,372 Hz	474 Hz
Hip-Hop / Rap	Power, 3,168 Hz	Peacefulness, 2,791 Hz	377 Hz
Latin	Tenderness, 3,054 Hz	Peacefulness, 2,742 Hz	312 Hz
Punk / Emo	Power, 3,310 Hz	Sadness, 3,037 Hz	273 Hz
Metal	Tension, 3,270 Hz	Wonder, 3,082 Hz	188 Hz

Metal is the most emotionally rigid format we measured. From its most tense record to its most wondering one, the entire range fits inside 188 Hz of spectral centroid. A metal record sounds like metal whatever it is feeling, because the format's signature does so much of the work that emotion barely moves the needle. Punk and emo are nearly as tight.

Scores and electronic are the opposite, with over a thousand hertz of usable emotional range. Which is why the same instrument set can read as menace or as calm depending purely on where the balance sits.

And peacefulness is the darkest emotion in twelve of the fifteen genres, which makes it the most portable signature in the study.

What I Will Not Claim

The mapping is a judgement call. Deciding hypnotic belongs to transcendence and psychedelic to wonder was my decision, not a measurement, and those two decisions matter enormously: rock carries psychedelic 1,275 times and hypnotic 49, while house is the reverse. So any apparent difference between transcendence and wonder by genre is substantially an artifact of where I filed two words. One rater, no reliability check. That is the weakest link in the chain.

Two tags were excluded deliberately. Warm and dark are among the most common in the set and they are tonal descriptions rather than emotions. If whoever assigned them was listening to frequency balance while doing it, including them would smuggle the circularity back in. Dropping them cost about 43,000 labelled observations and the structure did not change, which is the best evidence I have that it is real.

Genre and loudness were controlled separately, not jointly. I stratified by genre, and I compared loudness effects against spectral ones. I did not fit a single model holding both at once. A proper version of this study would.

The dark category is too small to trust, at 119 tracks. It is in the tables for completeness and should not be leaned on.

What This Changes in the Studio

I am not going to mix to a table. But three things have changed.

I will stop treating tempo as an emotional lever. If a record is not landing, the answer is in how hard it is pushed and where the middle sits, not in the grid.

When something needs to feel vast, I will look at the midrange before I look at the top. The scoop is the signature, and it is a bigger effect than the air is.

And when a record is meant to feel calm, I will check its brightness first, because that is the single most reliable relationship in 30,912 tracks.

The model gets fixed next. The wonder weight is wrong and comes out. The tempo weights come out. The transcendence weight stays but gets reduced, and gains a midrange term it never had. Then the whole thing goes back for another round against labels it has never seen.

---

To see where your own track sits, the Sonic Analyzer is free and takes about a minute. The earlier studies cover loudness by genre and tonal balance by genre. To talk through a record, book a call or check the rates.