Songs with stable, exposed pitch reveal a 432 Hz retune most clearly: dry lead vocals, sustained pads and strings, piano and acoustic intros, and long reverb tails. Sustained 808s and sub-bass show it as weight on full-range playback but are hard to judge on small speakers. Dense, distorted, drum-heavy sections hide it. Test with music you made, own, or are licensed to process, and keep anything that isn't yours private.

Here's something that took me a while to accept: the same conversion can be obvious on one song and nearly invisible on the next. People then decide either that 432 is magic or that it does nothing, when really they just picked a track that couldn't show them much. A third of a semitone needs something stable to push against.

So before you queue up a catalog, it's worth knowing which kinds of material actually reveal the change, which kinds hide it, and how to build a short test set that tells you something true about your own music.

Start with music you're allowed to change

This part is boring and important. The best test track is one you wrote, produced, bought, or licensed for exactly this kind of processing. The U.S. Copyright Office explains that a sound recording and the musical composition inside it are separate works with separate copyrights. So publishing a retuned version of someone else's record touches at least two sets of rights, and a frequency label doesn't change that. Private listening on a legally obtained copy is one thing. Uploading a "432 version" of a famous song to a public channel is another.

The good news is that your own songs are better test material anyway. You know them in your bones, which gives your ear a stable internal reference. Use famous records as references for characteristics (that exposed piano intro, that dry close vocal, that sub-heavy verse) and then find those same characteristics in music you control.

Why some material shows the change and some hides it

A classic retune multiplies every frequency by 0.981818. On a musical, logarithmic scale, that's a clean slide of the whole spectrum 31.77 cents to the left. Anything with clear, stable peaks, like the harmonics of a held note, visibly moves. Anything noise-like, like a cymbal wash, a snare crack, or a distorted wall, is roughly the same shape before and after, so there's very little for your ear to compare. There's nothing pitched to move.

Time matters too. Your ear needs a note to sit still for a moment to judge where it is. Long vowels, held chords, and slow decays give it that time. A sixteenth-note hat pattern doesn't. And familiarity matters most of all, because you're comparing against memory as much as against the other file.

Measuring it on synthetic archetypes

I wanted numbers rather than vibes, so this script builds seven synthetic 30-second archetypes (a vocal-like harmonic line with vibrato and a formant, piano-like decaying partials, a sustained detuned-saw pad, an E1 808, a reverberant chord stab, drums only, and a dense distorted mix), renders each through a classic 432 resample with ffmpeg, and measures how far the long-term spectrum moved on a musical (1/48-octave) grid. It also runs each through a rough phone-speaker filter to see what the retune costs on tiny drivers.

#!/usr/bin/env python3
"""reveal_score.py: which kinds of material expose a 440->432 retune?
Usage: python3 reveal_score.py
Needs: numpy, scipy, ffmpeg (with soxr) on PATH.

Builds seven SYNTHETIC 30-second archetypes, renders each through a classic
432 resample with ffmpeg, then measures:
  flatness  : median spectral flatness in dB (very negative = tonal,
              near 0 = noise-like)
  reveal dB : log-spectral distance between the 440 and 432 long-term
              spectra on a 1/48-octave grid, so a low note and a high note
              moving by the same cents count the same, floored at -60 dB
              re the loudest band. Big = the spectrum
              visibly moved; small = the change hides.
  phone dB  : level change through a phone-like 4th-order 150 Hz high-pass
              (negative = the 432 version loses more on small speakers)."""
import subprocess
import numpy as np
from scipy import signal

SR, D = 48000, 30.0
t = np.arange(int(SR * D)) / SR
rng = np.random.default_rng(5)

def harm(f, n, tilt=1.0, amp=1.0):
    ph = 2 * np.pi * (np.cumsum(f) / SR if np.ndim(f) else f * t)   # integrate frequency -> phase
    return sum(amp / h ** tilt * np.sin(h * ph + rng.uniform(0, 6.28)) for h in range(1, n + 1))

def make():
    out = {}
    vib = 1 + 0.004 * np.sin(2 * np.pi * 5.2 * t)
    v = harm(220.0 * vib, 24, 1.0)                      # A3 "vocal": harmonic stack...
    b, a = signal.iirpeak(700 / (SR / 2), 4); v = signal.lfilter(b, a, v) * 3 + v * 0.3   # ...with a formant
    out["dry vocal-like line"] = v
    env = np.exp(-1.2 * (t % 2.0))
    p = sum(np.sin(2 * np.pi * 261.63 * k * np.sqrt(1 + 0.0004 * k * k) * t) / k for k in range(1, 12))
    out["piano-like decays"] = p * env
    pad = sum(signal.sawtooth(2 * np.pi * f * (1 + d) * t) for f in (220.0, 261.63, 329.63) for d in (-0.002, 0, 0.002))
    out["sustained synth pad"] = signal.lfilter(*signal.butter(2, 3000 / (SR / 2)), pad)
    beat = (t * 70 / 60) % 1.0
    s808 = np.tanh(1.5 * np.sin(2 * np.pi * 41.2 * t) * np.exp(-1.0 * beat))   # E1 808, 70 BPM
    out["808 / sub-bass (E1)"] = s808
    ir = rng.standard_normal(int(SR * 2.5)) * np.exp(-np.arange(int(SR * 2.5)) / (SR * 0.6))
    chord = harm(196.0, 8) * ((t % 4.0) < 1.0)          # G3 chord stab every 4 s
    out["ambient tail"] = signal.fftconvolve(chord, ir)[: t.size]
    kick = np.sin(2 * np.pi * 50 * t) * np.exp(-25 * ((t * 2) % 1.0))
    noise = rng.standard_normal(t.size) * np.exp(-30 * ((t * 4) % 1.0))
    out["drums only"] = kick + 0.6 * noise
    gtr = np.clip(4 * (harm(110.0, 12) + harm(164.81, 12) + harm(220.0, 12)), -1, 1)   # clipped power chord
    wash = signal.lfilter([1], [1, -0.95], rng.standard_normal(t.size)) * 0.08         # constant cymbal/noise bed
    out["dense distorted mix"] = 0.6 * gtr + 1.2 * noise + wash + kick
    return {k: 0.5 * v / np.max(np.abs(v)) for k, v in out.items()}

def retune(x):
    p = subprocess.run(["ffmpeg", "-v", "error", "-f", "f32le", "-ar", str(SR), "-ac", "1", "-i", "-",
                        "-af", f"asetrate={SR*432/440:.6f},aresample={SR}:resampler=soxr",
                        "-f", "f32le", "-"], input=x.astype(np.float32).tobytes(), capture_output=True, check=True)
    return np.frombuffer(p.stdout, dtype=np.float32)

EDGES = 30.0 * 2 ** (np.arange(0, int(48 * np.log2(12000 / 30)) + 1) / 48)   # 1/48-octave grid
def ltas(x):
    f, P = signal.welch(x, SR, nperseg=65536)
    idx = np.digitize(f, EDGES)
    bands = np.array([P[idx == i].mean() if np.any(idx == i) else np.nan for i in range(1, len(EDGES))])
    return bands + 1e-14

def flatness_db(x):
    _, _, S = signal.stft(x, SR, nperseg=4096)
    P = np.abs(S[1:]) ** 2 + 1e-14
    flat = np.exp(np.mean(np.log(P), axis=0)) / np.mean(P, axis=0)
    return 10 * float(np.log10(np.median(flat)))

phone = signal.butter(4, 150 / (SR / 2), "highpass", output="sos")
lvl = lambda x: 10 * np.log10(np.mean(x ** 2) + 1e-20)

rows = []
for name, x in make().items():
    y = retune(x)[: x.size]
    A, B = ltas(x), ltas(y)
    floor = 1e-6 * np.nanmax(A)                    # -60 dB re peak: a stand-in for the noise floor
    ok = ~np.isnan(A) & ~np.isnan(B) & (np.maximum(A, B) > floor)
    A, B = np.maximum(A[ok], floor), np.maximum(B[ok], floor)
    lsd = float(np.sqrt(np.mean((10 * np.log10(A / B)) ** 2)))
    ph = (lvl(signal.sosfilt(phone, y)) - lvl(y)) - (lvl(signal.sosfilt(phone, x)) - lvl(x))
    rows.append((name, flatness_db(x), lsd, ph))

print("SYNTHETIC archetypes, classic 440->432 resample (ratio 0.981818, -31.77 cents)")
print(f"{'material':24s} {'flatness':>9s} {'reveal dB':>10s} {'phone dB':>9s}")
for name, ton, lsd, ph in sorted(rows, key=lambda r: -r[2]):
    print(f"{name:24s} {ton:9.1f} {lsd:10.2f} {ph:+9.2f}")
print(f"E1 fundamental: 41.20 Hz -> {41.2*432/440:.2f} Hz;  A4: 440 -> 432 Hz;  1 kHz partial -> {1000*432/440:.1f} Hz")

Real output from this machine. Every signal here is synthetic and generated by the script; no commercial recording was used.

SYNTHETIC archetypes, classic 440->432 resample (ratio 0.981818, -31.77 cents)
material                  flatness  reveal dB  phone dB
dry vocal-like line          -80.2      22.50     -0.00
sustained synth pad          -46.3      15.83     -0.01
piano-like decays            -74.9       7.58     -0.01
ambient tail                 -41.4       7.26     -0.04
dense distorted mix          -18.3       5.64     -0.05
808 / sub-bass (E1)          -93.4       3.21     -0.40
drums only                    -2.8       2.06     -0.13
E1 fundamental: 41.20 Hz -> 40.45 Hz;  A4: 440 -> 432 Hz;  1 kHz partial -> 981.8 Hz

Reading the numbers

The "reveal" column is the useful one. It's the average spectral difference, in dB, between the 440 and 432 versions once you line them up on a musical grid. Bigger means the change is more visible.

The vocal-like line tops the list at 22.5 dB, with the sustained pad close behind at 15.8. Both are dense with stable harmonics and leave quiet gaps between them, so a 31.77-cent slide moves energy into places that were nearly silent before. Piano decays and the ambient tail sit in the middle around 7 dB. The dense distorted mix drops to 5.6 even though there's a chord inside it, because clipping and a constant noise bed fill the gaps the chord would otherwise move through. Drums alone bottom out near 2 dB. The 808 lands low at 3.2, for a reason that's worth its own section.

Two honest caveats. These are synthetic signals, deliberately pure, so real records will compress the range. And a spectral measurement isn't a listening test; it tells you where to look, and a blind, level-matched listening test tells you what you actually hear.

Vocals: the most revealing, and the riskiest

A dry lead vocal is the single best test you have. You know exactly where that voice sits, and the harmonics are stable through every held vowel. Listen to verse lines and sustained notes, not ad-libs over a busy beat.

This is also where 432 can lose. A lower reference can suit one singer beautifully, adding chest and weight, and make another sound slightly dull or tired. Formants move along with pitch in a classic retune, so the whole voice shifts a little darker. If the intro gets better and the chorus loses energy, that isn't a failed test. That's production information.

Pads, strings, organs, and piano

Sustained synths, strings, and organ chords are the second-best test, and slow intros are gold. Piano and acoustic guitar sit just behind, because the decays are long and the partials are clear but the notes do eventually die away. Listen for weight and relationship between notes rather than automatic improvement. You'll often hear the lower version as a little softer or less urgent; whether that's "warmer" or "flatter" is your call.

808s and sub-bass: felt more than seen

The 808 result looks odd at first, so here's what's going on. An E1 808 at 41.2 Hz moves to about 40.45 Hz. That's the same 31.77 cents as every other note, but it's only 0.75 Hz, which is hard for both a spectrum analyser and an ear to resolve at that depth. What you notice on a capable system is weight: the note sits lower and the slightly slower tail of a classic resample makes it feel heavier. Slow trap, chopped-and-screwed, ambient, and vaporous R&B often wear this well.

The phone column is the warning. Through a phone-like filter, the 432 version of the 808 lost an extra 0.4 dB compared with the 440 one, because a fundamental that was already near the speaker's lower limit moved a little further out of reach. Small, but real. Judge bass-led records where your audience actually listens, and if the 808 is already fighting the kick, sort that out first; Your kick and 808 are arguing in the same chair covers it. If phones are your main audience, also check what mono fold-down does to the low end and the width while you're there.

Dense mixes and drums: where it hides

Loud, distorted, percussion-led choruses are where most people decide "I can't hear a difference." The numbers agree. There's little stable pitch for the ear to track, and what's there is buried under noise. That's useful to know, not disappointing. If your catalog is mostly dense, aggressive records, a 432 alternate may simply not be worth managing. If it's mostly sparse and melodic, it might be.

What reveals it, at a glance

Check the song's real starting pitch

One more trap before you score anything. Not every record starts at A=440. Older tape transfers can run slightly fast or slow, plenty of producers detune samples or whole beats for colour, some orchestral recordings sit at 442 or 443, and a file you found labeled "432" may already have been converted once. Push any of those through a 440-to-432 conversion and you get a different distance than you think, sometimes enough to change the result.

A tuner plugin on a sustained note, or a spectrum reading of a held bass note, usually tells you within a few cents whether the source sits where you assume. If it doesn't, note the real reference and convert from that, or leave the song out of the test set. A clean test needs a known starting point as much as it needs a revealing passage.

Build a ten-track test set

Pick ten songs deliberately, from music you own or control: one intimate vocal, one harmony-heavy chorus, one acoustic instrument, one piano-led piece, one sustained synth track, one bass-led record, one percussion-heavy mix, one dense distorted mix, one ambient piece, and one song whose pace already feels slightly rushed. Before converting anything, mark a 15 to 30 second region in each that shows its defining character.

Then score each on a 1 to 5 scale for pitch weight, brightness, pace, vocal fit, low-end translation, and overall preference, ideally blind. The pattern across ten songs matters far more than your favourite single result. You might find 432 consistently suits spacious, bass-led, melodic material and drains urgency from aggressive tracks. That selective answer is more valuable than converting everything because a label is fashionable.

Only render the winners

Once a song wins a fair comparison, give it a proper full-length render, measure it, and deliver it as a clearly labeled alternate. 432 Hz mastering: LUFS and true peak after a retune covers that delivery step. The rest stay as they are, and that's fine. The point of a test set isn't to prove 432 right or wrong; it's to find out which of your records it actually serves.

What songs are best for hearing the difference between 432 and 440 Hz?

Songs with stable, exposed pitch: a dry lead vocal, a slow piano or acoustic intro, sustained pads, strings, or organ. Songs you know well work best because you already carry their pitch in memory.

Why can't I hear a 432 retune on some tracks?

Dense, distorted, percussion-heavy sections have little stable pitch for your ear to track, so a 31.77-cent move hides inside the noise. That's normal and doesn't mean the conversion failed.

Does 432 Hz make 808s and bass sound heavier?

Every note moves down by the same ratio, so an E1 808 at 41.2 Hz lands near 40.45 Hz. On a full-range system that can feel weightier. On small speakers that can't reproduce it, you lose a little more of the fundamental.

Can I retune and upload a famous song in 432 Hz?

Not without permission. A sound recording and the musical composition are separate copyrights, and distributing an altered version of someone else's recording needs the right permissions. Test privately on music you own, or use your own catalog.

How many songs should I test before converting a catalog?

Ten deliberately different tracks is a good start: vocals, piano or acoustic, sustained synths, bass-led, dense, and ambient. Render only the songs that win a blind comparison.