Listening and comparing

Two versions of the same song. One of them has been "improved". Listen to both, as many times as you like, and pick the one you prefer. The names are hidden until you answer.

Now press Reveal. The only difference is that one version is 1.5 dB louder. No EQ, no compression, no magic.

If you picked the louder one and described it as "fuller", "clearer" or "more present", you're in very good company. Many people do, and most of them don't hear it as louder at all. They hear it as better. If you heard no reliable difference, that's an honest answer too, and a useful one.

Your ears are honest, but easily flattered

A little extra level makes almost everything sound more impressive. You heard why in Sound: a louder playback brings the deep bass and the top end up into the range where your hearing is sharpest. So the louder version really does sound fuller. Your ears aren't lying. They're just answering a different question from the one you asked.

This is the trap behind most bad mixing decisions. Many effects add a little level as a side effect: a compressor with makeup gain, a saturator, an EQ boost, an exciter. Switch one on, it gets louder, it sounds better, you keep it. Do that twenty times and you've built a chain that mostly makes things louder.

The fix is simple to say: before you judge a change, take away every advantage that isn't the change.

Ask a smaller question first

"Is it better?" bundles together tone, dynamics, loudness, style and mood. It's too big to answer. Before you touch anything, finish one sentence:

"I want this change to make ___ easier to hear."

Fill the blank with something concrete: the first hit of each bass note, the last word of the chorus, the snare's crack. "Rougher" and "darker" are fine goals too, as long as you'd recognise success. Now you have something to listen for, rather than a vague feeling to chase.

Remove the level advantage

Here's a compressor on a drum loop. Before you touch it, notice the reading that compares the level with the compressor on and Dry.

With the makeup gain up, the compressed version sounds huge. Press Match and the lab turns the makeup down until both sides are equally loud. Switch again. Now you're hearing only what the compressor actually did: the attacks of the hits, the swell of the room between them, the way the snare tail stands up. Maybe you still love it. Maybe it just sounds squashed. Either way, now it's a real decision.

A few habits make every comparison fairer:

• Same passage, same start point. A verse against a chorus tests the arrangement, not the plugin.
• Turn the louder one down, not the quieter one up. You keep headroom and your ears stay rested.
• Don't touch the volume knob between versions. Make the compensation inside the comparison so you can repeat it.
• Listen in context when the goal involves other parts. A vocal that sounds great alone might bury the guitar.

Hide the names

Labels steer your ears before the sound starts. "Vintage", "analog", "expensive", "the new mix", even "my favourite preset" all prime you to hear what you expect. Wine tasters, violinists and audio engineers have all been caught by this one. Hiding the names lets your ears work alone.

Without a lab, ask a friend to play two versions to you as X and Y, in a random order, and to write down which is which. Working alone, render both, give them neutral names and let your music player shuffle them. Make sure nothing gives the answer away: a filename, a waveform's size, a fader you can see.

The formal version of this is called an ABX test: you hear A, you hear B, then you hear X and say which one it is. If you can't pick X better than a coin flip over a decent number of tries, the difference may be smaller than you thought.

Hearing a difference isn't the same as liking it

There are two separate questions, and they deserve separate answers:

1. Can I reliably tell these apart?
2. Which one serves the song?

You can hear a difference clearly and have no preference. You can prefer one version in the chorus and the other in the verse. You can fail to tell them apart and still decide to keep the simpler one, because it uses less CPU and has fewer things to go wrong. None of these are contradictions.

Write down your reason every time. "X has a longer tail" is a finding. "X wins" is a mood. Ten quick trials won't prove anything to a scientist, but they'll show you whether you're consistent, and that's what matters at 2 a.m. with a deadline.

Judging a model's output

The same discipline matters even more when the change comes from a model: a stem splitter, a note reader, a denoiser, a sound generator. Model tools usually arrive with a score attached, and the scores are worth knowing, as long as you know what they measure.

• Note accuracy. A note-reading model is scored by how many of the played notes it found, and how many extra notes it invented. A note usually counts as "found" if it starts within a small window of the real one, often about 50 milliseconds. That's the default in mir_eval, the open-source library many researchers score with. So a high score can still hide a rhythm that feels slightly off, and a wrong octave can count as a miss even when the tune is plainly right. What an audio model learns shows the kinds of mistakes behind those numbers.
• Separation quality. A stem splitter is scored by comparing its estimate of each part with the original track, in decibels, a score the field calls signal-to-distortion ratio: higher means less of everything else got in. The number blends leakage, missing sound and watery artefacts into one figure, so two stems with the same score can fail in very different ways. Stem separation explains why you can't simply un-mix a song.
• Distance scores for generated audio. A generator is often scored by how close the overall character of a big batch of its clips is to a batch of real recordings, with a Fréchet-style distance such as Fréchet Audio Distance. It's a score for the batch, not for any one clip, and it says nothing about whether a sound fits your track.

Treat these numbers as rough guides for choosing which tools to try, not as the verdict. They're measured on someone else's test recordings, at someone else's settings, for someone else's idea of "correct". The question that matters is the one from the top of this guide, asked of your own material: does this change make the thing I care about easier to hear? Match the level, hide the names, and listen to a passage of your song. A tool with a lower score that fails somewhere you don't care about can be the better choice.

Where careful comparison goes wrong

• Matching on the wrong span. A compressor changes the loud parts more than the quiet ones. Match on the chorus and the verse may be way off. Listen to the full range of material the setting will process.
• Endless switching. After the fortieth A/B, every version sounds the same and slightly wrong. Take a break, then listen once through without touching anything.
• Testing ten things at once. One source, one goal, one change. Broaden only once you know why.

Same idea, different place: hiding the label is the same move as Ear training, where you guess an interval before seeing its name. In both cases, you're keeping what you expect from deciding what you hear.

Make it yours

Pick one effect from a project of your own, one you've had switched on so long you've stopped questioning it. Finish the sentence: "This makes ___ easier to hear." Match its level against bypass, hide which is which, and run ten short trials. Write down your guesses and your preferences separately.

Then listen to a different section of the song. Does the benefit survive?

Keep or remove it, and write one line explaining why, plus the gain you used to match. That line is worth more than the preset. Months from now, it will tell you whether the effect was doing a job or just turning things up.

Quick check

1. You add a saturator to a bass and it instantly sounds richer. The meter shows it's 2 dB louder. Have you learned anything about the saturator yet? Not much. Turn it down by 2 dB, compare again on the same passage, and then decide. The richness might survive, or it might have been the level all along.
2. In a blind test you correctly pick the processed version eight times out of ten, but you prefer the original. Is the test broken? No. You can clearly hear a difference and still not like it. Detecting and preferring are separate questions.
3. You match a compressed vocal to the dry one on the loud chorus, and now the quiet verse sounds much louder than dry. What happened? The compressor changed loud and quiet parts by different amounts, so one match can't fit every moment. Check both sections and decide which behaviour the song needs.

Next steps

• Go deeper: Compression. The effect most likely to fool you this way, because it changes level by design.
• Use it: Finishing and references. The same discipline, applied to a whole mix next to a song you love.