Finagle My Song · Blog
I ran 43 real mixes through my own analyzer. It said 42 were bass-heavy.
I built a free mix analyzer. You upload a track and it tells you what's going on with your low end, your top end, your loudness and your stereo image.
This week I finally did what I should have done at the start. I fed it 43 mixes from the Cambridge-MT library, the free multitrack collection that engineers use to practise and teach mixing. Each multitrack there comes with the band's finished mix, and those are what I tested.
My analyzer told 42 of them they were bass-heavy and would sound muddy on earbuds. It told all 43 that the top end was thin and needed a presence boost. Not one of them got "balanced low end".
So the analyzer was wrong, 43 times over.
Where it went wrong
Most mix analyzers, mine included, split the spectrum into bands and tell you what percentage of the energy sits in each one. Sub, bass, low-mids, high-mids, treble.
The mistake was in what counted as normal. My thresholds treated those bands as if a healthy mix spreads its energy roughly evenly across them. Anything over 55% in the sub and bass got flagged as heavy.
Real mixes don't work like that. Music naturally carries most of its energy down low. The kick, the bass and the fundamentals of most instruments all live there. Across the 43 mixes, the median share of energy below 300 Hz was 86%. The heaviest was 93%. Even the acoustic and a cappella tracks sat above 50%.
So 86% bass is simply what a finished mix looks like.
The same mistake hid a real problem
This is the part that actually cost me.
If bass takes most of the energy, then everything above it looks tiny as a percentage, even when it isn't. Treble on a typical mix measures well under 1% of the whole spectrum. So a percentage readout makes nearly every mix look dull.
I had a master of my own that was genuinely harsh. It measured 3.0% treble. A clean master of another song measured 2.9%. On paper they were the same, and the AI in my own tool read the harsh one as dull and suggested adding more top end. That's the one thing it shouldn't have done.
The fix is to measure brightness against the musical range only, leaving the bass out of the sum. Looked at that way, the 2 to 5 kHz region takes about 9% of the energy above 200 Hz in a typical mix, and 90% of mixes stay under 18%. My harsh master sat at 33%. Suddenly the problem was obvious, and the tool now cuts there instead of boosting.
"Heavy processing detected"
My analyzer also had a health check that warned when a file looked over-processed. It flagged 6 of the 43 mixes.
Every one of the six was simply loud, somewhere between -6 and -8 LUFS. The check was counting loudness as damage, and a loud master can be perfectly healthy. Plenty of great records are loud on purpose.
Real damage leaves a different fingerprint. When a signal clips, the tops of the waveform get chopped flat. I deliberately wrecked copies of five mixes, hard-clipped and brickwall-crushed, and compared them with the real ones. Flat-topped samples showed up in every wrecked file and in none of the 47 real files I tested (the 43 plus four of my own). That's what the check looks for now.
Stereo width
One more. My AI assistant told me a mix with a stereo correlation of 0.2 was "dangerously close to out-of-phase" and might cancel itself out in mono.
It won't. Correlation runs from +1 (identical channels) to -1 (one channel is the other flipped). Actual cancellation happens below zero. At 0.2 you've got a very wide mix, and the median across the 43 was 0.75. The lowest, at 0.15, was an a cappella recording. None went below zero.
A wide mix is worth a quick mono check, and that's all.
What to take from this
If you use any mix analyzer, including mine, a few things are worth knowing.
Percentages mean nothing without a reference. "40% bass" tells you nothing until you know what a finished mix in your genre measures. The most useful thing any analyzer can do is compare your track against a reference you trust, in the same tool, on the same scale.
Judge brightness with the bass taken out. If your mix is harsh, a whole-spectrum readout can hide it completely. Trust your ears here, or a measurement that ignores the low end.
Loud and damaged aren't the same thing. If a tool calls your master over-processed, ask what it's actually measuring. Clipping is damage. Loudness is a choice.
Low correlation isn't phase cancellation. Worry when it goes below zero. Above that, check your mix in mono and trust what you hear.
What changed
I've recalibrated the free analysis in Finagle My Song against those 43 mixes. Bass is judged against what real mixes carry. Brightness is measured with the bass taken out. The health check looks for clipping, not loudness. The AI can no longer add top end to a track that's already bright, or suggest widening a mix that's already wide.
I'd rather tell you I got this wrong than keep telling you your mix is bass-heavy when it isn't. Upload a track and see what it says now. It's free and there's no account.