Diagnosis

The absolute spectrum of a single unknown voice cannot be diagnosed.

Every voice has peaks and dips of exactly the size an injected filter has. Formants are ±10 dB. A speaker’s fundamental sits anywhere from 85 to 255 Hz. Sentences decline 5–6 dB from start to end as a matter of ordinary prosody. Look at one spectrum on its own and you will find “defects” in all of it, and the ones you find will be the speaker.

So diagnosis is always comparison. The whole method is choosing the right thing to compare against.

Compare against something inside the same file

Ranked by how much they can tell you. Prefer the highest one available.

1. The clean original, if it exists

If the undamaged take is on disk, this is the whole job — measure both, subtract, and the difference is the defect. Nothing below is as good. Look for it before anything else.

2. The pauses

The strongest reference that lives inside a single file. Speech stops; whatever is still there in the gap is not the voice.

What it answers: “was something added?”

Anything audible in the pauses is additive — hum, rumble, hiss, room tone. It was laid on top, so it can be subtracted, and this is a reliable positive finding.

What it does NOT answer: “was something filtered?” — and getting this backwards is how the method produces a confident wrong answer.

A filter multiplies. Applied to a file whose gaps already sit at the quantisation floor, it leaves them at the quantisation floor: near-silence times anything is still near-silence. So the pause carries no trace of it. Measured on one take with a −9 dB shelf above 2.5 kHz applied to the whole file:

The defect is a clear 5 dB in the speech and exactly zero in the pause.

So: never use a null result from the pause spectrum to rule out EQ. A run that did exactly that — measured the pause, found it smooth, and concluded “static EQ of any type or Q is ruled out” — went on to treat an inaudible −72 dBFS rumble as the defect and shipped a high-pass for a file whose actual problem was that it had no top end.

The pause spectrum is a transfer function only when the gaps carry a real recorded noise floor that passed through the same filter. A room-tone bed does; a digitally clean take does not. Check which you have before trusting it: if the gaps are within a few dB of the quantisation floor, this reference can find additive content and nothing else.

3. The speech’s own tilt, for a suspected filter

When the pause cannot see a filter (above), the only thing left carrying it is the speech. Read the tilt across a few 1/3-octave bands rather than any single one — 1k / 3.2k / 5k / 7k is enough to see a shelf:

Speech falls away steadily above about 1 kHz, so a downward slope is expected; what you are looking for is a slope that keeps steepening, or a step. In the table above, −8.1 dB from 1 k to 5 k is an ordinary voice and −13.1 dB is the same voice with 9 dB taken off the top.

This is a candidate, not a verdict. Where the ordinary slope ends and a defect begins is speaker-dependent, and you have no baseline for this speaker. Say what you measured and what it would mean, and let somebody hear it.

4. The file against itself over time

For anything level-related, compare each passage to the track’s own median rather than to a target. That is what levellingResult does, and it is why an already even track comes back untouched.

Do not compare against a different voice

Both wrong answers in the evaluation that produced this page came from an external reference, and both were argued rigorously from bad ground:

• A published average spectrum (LTASS and friends). One run concluded “+10 dB above 7 kHz, split-half stable, gating-independent” on a file whose actual defect was +6.6 dB at 200 Hz. Its supporting claim — 10 kHz sitting 6.2 dB above 6.3 kHz — measured 0.6 dB on re-check, and measured the same in the clean original. Published curves are mixed-sex, mixed-corpus, and mixed-microphone; the gap between them and any one speaker is larger than most defects.
• A synthesised control voice (say, a TTS take, another narrator). One run generated a control this way, found the spectrum “normal”, and missed a −6.9 dB shelf. Two speakers differ by more than 7 dB across the top octaves as a matter of course, so a cross-voice comparison cannot resolve a defect that size.

If neither the original nor usable pauses exist — continuous speech, or gaps that are digital silence and so carry no channel — then a static tonal defect is genuinely under-determined.

Report that. It is a finding, not a failure to find one, and it is the correct answer rather than the fallback when the better methods are unavailable. Give the author the two or three readings that fit and ask which they hear; they can listen, and that one sentence from them collapses the whole problem.

This is the point where a capable agent goes wrong. Told a thing is under-determined, the instinct is to invent a cleverer measurement and escape it — and something will always be found, because a single voice’s spectrum is full of peaks and valleys that survive any amount of statistical rigour. An elaborate novel method reaching a confident conclusion, on a file where the two reliable references were both unavailable, is the signature of this failure, not evidence against it. If you notice yourself building one, stop and report the ambiguity instead.

Recipes

Compare loudness from the bytes the listener actually hears

Do not call two clips equally loud because their Studio faders, waveform peaks, or cached asset metadata match. Those are controls and proxies, not a loudness measurement. Resolve the exact URLs used by preview/render, download or inspect those exact served bytes, and measure each decoded stream with FFmpeg’s ebur128 filter. Compare the integrated LUFS values.

For a target loudness, the required move is:

When both clips are local authored <audio> elements with stable ids, use the CLI instead of transcribing that arithmetic by hand:

The first command is a dry run. The second writes only the target’s data-volume, after accounting for both existing gains and refusing a boost that would clip or exceed Studio’s ceiling. Always choose the reference from the author’s stated intent; the command does not guess which clip should define the mix.

Studio’s clip-gain fader uses 0 dB / linear gain 1 at its physical midpoint and provides up to +12 dB on the upper half. After changing gain, measure the served preview/render bytes again. If a listener still hears a mismatch, trust the report and first verify the asset URL and bytes are current; do not explain it away with matching peaks or a stale proxy measurement.

All verified with ffmpeg 8.1.1. -hide_banner keeps the output readable; volumedetect prints to stderr, so do not silence it with -v error.

Band energy, in proportional bands

Use proportional bandwidths or the numbers lie. A fixed 2000 Hz-wide band at 10 kHz collects more energy than a 1200 Hz-wide band at 6.3 kHz for no reason but its width, which manufactures a high-frequency excess that is not there. One third of an octave is f × 0.2316.

Read them as a shape across 100 / 200 / 400 / 1k / 3.2k / 7k, and read the shape against a reference from the list above — never on its own.

The noise floor, and what is in it

-inf means digital silence in the gaps: no additive noise, so rumble, hiss and room tone are all ruled out in one command. A real number is the level of whatever is sitting under the voice. To see its shape, cut a pause out with -ss/-t and run the band recipe on that slice alone.

Level over time

LRA under ~3 LU is even. Then window it, because LRA hides a single sagging passage:

A 4–6 dB spread across windows is normal speech, not a defect — sentences decline as they end. Injected unevenness looks like 12 dB or more. Levelling a track that only has declination flattens the prosody and is heard as robotic.

Pitch, before blaming the low end

A voice has no energy below its own fundamental, so a “missing” 100 Hz on a speaker whose F0 is 210 Hz is the speaker, not a rolloff.

The same fact runs the other way, and that direction is the trap: a boost near the fundamental is indistinguishable from that voice being naturally chesty. Both look like energy at F0, because both are.

So the rule is symmetric, and the dangerous half is the second one:

• Do not call a peak at F0 a defect on its own evidence.
• Do not dismiss one either. “The peak is at 200 Hz, F0 is 185 Hz, therefore it is the fundamental” is not a diagnosis — it is the same observation restated, and it discards the one candidate most likely to be real. Boominess is excess energy at the bottom of a voice; that is what the word means.

What you can do is measure how much, against the same file’s midrange:

In an ordinary take these land within a couple of dB of each other. A low band sitting more than about 4 dB above the 1 kHz band is a strong boom or mud candidate. Measured across one voice damaged several ways: undamaged +0.9, harsh +0.6, dull +2.0; boomy +6.7, muddy +5.8. Treat the figure as indicative rather than a threshold — it is one speaker — but the separation is wide, and a reading up at +6 is worth raising even when you cannot explain it.

It still cannot tell you whether a filter did that or the speaker did, so report it as a candidate. That is the whole answer here: measure it, name it, hand the choice to somebody who can hear it.

Then, and only then, the symptom table

Measurement gives you the band and the kind. SKILL.md‘s table and presets.md‘s fuller one turn that into a fix. Going the other way round — picking a plausible fix and finding evidence for it — is how both wrong answers in the evaluation happened, and both were long, careful and confident.

One habit that catches it: before applying anything, state what you would expect to measure if you are wrong, and check that too.