Mantra Field Books · No. 7
The physics, the room, the microphone, the ear — and which audiophile claims survive a level-matched blind test.
Almost everything that goes wrong with sound goes wrong before anybody touches a fader. The room is wrong, the gain is wrong, the microphone is in the wrong place, or the person listening has been listening for six hours and can no longer hear.
This book is about those things. It is written by a working video editor who mixes for broadcast, for somebody who wants to understand what is physically happening rather than collect techniques.
There is a companion guide in this series — Look at one frame. Hear what belongs in it. — on AI-assisted sound design, written out of the making of The Brain Brake. It covers the four kinds of sound, the prompt as specification, and test discipline for a generative pipeline.
It does not cover acoustics, and this book does not repeat it. The two are meant to be read together: that one is about deciding what a scene should sound like; this one is about the physics and the equipment that determine whether you can hear it.
One promise about tone. Chapter twelve is about hi-fi claims and measurement, and it will be unpopular in places. I have tried to separate three things carefully — what has been tested and found audible, what has been tested and not found audible, and what has simply never been properly tested. The third category is larger than either camp admits, and pretending otherwise is how this argument stays stupid for another thirty years.
Chapter one
Pressure, frequency, wavelength, phase — and why the numbers matter in a room.
Sound is a pressure wave in a medium. Air molecules are pushed together and pulled apart, and that alternation propagates. Nothing moves from the source to your ear except the pattern.
Sound travels at roughly 343 metres per second in air at 20°C. Wavelength is that speed divided by frequency, and once you have those figures in your head, a great deal of otherwise mysterious behaviour becomes obvious.
| Frequency | Wavelength | What that implies |
|---|---|---|
| 20 Hz | 17.2 m | Longer than your building. It does not care about your walls. |
| 50 Hz | 6.9 m | Bigger than most rooms. This is why bass is the hard problem. |
| 100 Hz | 3.4 m | About one room dimension. Chapter four is about what happens next. |
| 1 kHz | 34 cm | Roughly a head width. Where localisation cues change character. |
| 10 kHz | 3.4 cm | Small. Absorbed by almost anything soft, and highly directional. |
This single table explains why a duvet on the wall does nothing for bass, why a subwoofer's position matters more than its quality, and why moving your head twenty centimetres changes what you hear.
When two copies of a signal arrive at slightly different times, some frequencies add and others cancel, producing a series of notches across the spectrum — comb filtering. It sounds hollow, phasey, thin, and it is the single most common cause of a recording that is technically clean and somehow wrong.
It happens when you use two microphones on one source, when a direct sound meets its own reflection off a desk, and when a monitor's output meets its own bounce off the wall behind you. It is not an equipment fault. It is arithmetic.
Chapter two
dBSPL, dBFS, dBu — why the same number means four things.
The decibel confuses people because it is not a unit. It is a ratio, expressed logarithmically, and it means nothing until you say what it is a ratio to.
| Notation | Referenced to | Used for |
|---|---|---|
| dB SPL | The threshold of hearing (20 µPa) | Actual acoustic loudness in a room. What a sound level meter reads. |
| dBFS | Digital full scale, the maximum representable value | Levels inside a DAW. Always negative; 0 is the ceiling, not a target. |
| dBu | 0.775 V | Analogue line level. Professional nominal is +4 dBu. |
| LUFS | Full scale, K-weighted, averaged over time | Perceived loudness for delivery. Chapter eleven. |
A measurement in dB(A) has been filtered to approximate the ear's reduced sensitivity to low frequencies. It is the standard for occupational noise limits and it is genuinely useful. But it was derived from equal-loudness behaviour at low levels, and it systematically under-reports bass energy. A loud club measured in dB(A) reads lower than what your body is receiving. Where the sound is bass-heavy, dB(C) or unweighted tells you more.
Chapter three
Equal loudness, masking, localisation, and the damage that does not come back.
The cochlea is a fluid-filled spiral whose internal membrane varies in stiffness along its length, so different positions resonate at different frequencies. Hair cells at each position report their activity. It is, functionally, a mechanical spectrum analyser with a nerve attached — and it is where the useful facts start.
Sensitivity is not flat. The ear is most sensitive around 2–5 kHz — the speech region — and much less sensitive at the extremes. Critically, the shape changes with level: at low volume you lose bass and treble disproportionately, which is why a mix that sounds full at conversation level can be thin when played loud, and why quiet mixing produces bass-heavy results. The ISO 226 equal-loudness contours are the formal version of this.
Practical consequence: mix at a consistent level, and check at two or three others. The absolute level matters less than its consistency, because your reference is your own memory of how things sound at that level.
A loud sound hides a quieter one nearby in frequency. The cochlea's resolution is limited — energy in one region raises the detection threshold across a band around it — and this is why a mix gets muddy: instruments are not too loud, they are occupying the same region and hiding each other. It is also the entire basis of MP3 and AAC, which discard what masking predicts you cannot hear.
The mixing implication is the most useful single idea in chapter nine: cutting the masker usually works better than boosting the masked.
Below roughly 1 kHz the head is small compared to the wavelength, so the brain uses arrival-time differences between the ears. Above it, wavelengths are short enough that the head casts an acoustic shadow, so level differences dominate. Front-back and height come from the filtering imposed by the shape of your outer ear, which is personal — the reason generic binaural audio works well for some people and not others.
Hair cells do not regenerate. Damage accumulates and does not reverse, and the first loss is usually in the range that matters most for consonants — so speech becomes hard to follow in noise long before anything sounds obviously wrong.
The European framework (Directive 2003/10/EC) sets a lower action value at 80 dB(A) daily exposure, an upper action value at 85 dB(A), and an exposure limit of 87 dB(A). Because 3 dB doubles the power, safe duration roughly halves for every 3 dB above the reference — meaning eight hours at 85 becomes four at 88 and two at 91.
If you are a percussionist, a DJ, or an editor on headphones all day — all three, in my case — you are in an occupational risk group and moulded filtered earplugs are the cheapest professional insurance available. Tinnitus arrives without warning and does not leave.
Chapter four
Modes, reflections, RT60, comb filtering, and the small-room problem.
You are not listening to your speakers. You are listening to your speakers plus your room, and in a domestic-sized space the room usually contributes more error than every other component combined.
Between two parallel surfaces, a standing wave forms at the frequency whose half-wavelength fits the distance — and at every multiple of it. The lowest axial mode is roughly 172 divided by the dimension in metres. A 3.4 m wall gives a mode near 50 Hz; a 5 m wall near 34 Hz.
At a mode's peak positions the frequency is exaggerated by many decibels; at its nulls it can nearly vanish. Move a metre and the bass changes completely. This is why bass sounds different on your sofa and at your desk, and why a mix made in an untreated room translates badly.
Two consequences. First, square rooms are the worst possible shape, because dimensions that match stack their modes at the same frequencies. Second, mode positions depend on the room, so no amount of speaker quality fixes them.
RT60 is the time for sound to decay by 60 dB. A cathedral runs to seconds; a small control room wants something in the region of 0.2–0.4 s, and — more important than the number — reasonably even across frequency. A room that soaks up treble but not bass measures short and sounds boomy and dead at once, which is what happens when somebody staples foam to everything.
The first reflection points — where sound from each speaker bounces off the nearest side wall, ceiling, floor and desk on its way to you — arrive within a few milliseconds of the direct sound and comb-filter with it, smearing the stereo image.
Below roughly 300 Hz, a domestic room does not behave like a diffuse acoustic space. It behaves like a resonant box with a handful of strong modes. There is no treatment that removes this, only measures that reduce it. Anybody working in a spare bedroom is working around a problem they cannot solve, and knowing that is the difference between adjusting sensibly and buying monitors that will not help.
Chapter five
In order of effect, with what each part costs.
Ranked by how much difference each makes per unit of money and effort. Do them in this order.
A porous absorber works when it is a meaningful fraction of the wavelength thick. Thin foam — two or three centimetres — absorbs high frequencies effectively and does essentially nothing below a few hundred hertz. Cover a room in it and you remove the treble while leaving every bass problem exactly where it was, producing a room that is simultaneously dull and boomy.
The physics is unforgiving and simple: to absorb low frequencies you need depth, either of material or of trapped air behind it. Mineral wool in a wooden frame costs a fraction of branded foam and works, because it is thick.
One measurement note: a calibrated microphone and free software will show you your room's frequency response and decay in an afternoon. Guessing is optional and expensive.
Chapter six
Transducers, patterns, proximity, and the stereo arrays worth knowing.
Every directional microphone boosts low frequencies as the source gets closer. This is why radio voices sound authoritative and why an inexperienced singer moving back and forth produces a level and tonal mess. It is a tool once you know it exists, and it does not occur on omnidirectional mics.
When using multiple microphones on one source, place each at least three times as far from the others as it is from its own source. This keeps the bleed low enough that the comb filtering of chapter one stays inaudible. It is a rule of thumb, not a law, and ignoring it is the usual cause of a multi-mic recording sounding hollow.
| Array | Arrangement | Character |
|---|---|---|
| XY | Two cardioids coincident, ~90° | Mono-safe, reliable, modest width |
| ORTF | Two cardioids, 17 cm apart at 110° | Approximates the head. Excellent general choice. |
| Spaced pair (AB) | Two omnis, well apart | Wide and spacious; centre can wander |
| Mid-Side | Cardioid forward + figure-8 sideways | Width adjustable after recording. Perfectly mono-compatible. |
| Blumlein | Two figure-8s crossed at 90° | Superb imaging, and it records the room behind you |
For a frame drum or any hand percussion, ORTF at a metre or so is a reliable starting point, and a spot mic close to the shell will give you body to blend in. Close-miking a frame drum alone tends to capture the mechanics of the hand rather than the instrument.
Chapter seven
Gain staging, converters, sample rate, bit depth, dither, latency.
The principle is one sentence: get the level right as early as possible, because noise added early is amplified by everything after it, and distortion added early cannot be removed.
In practice, with 24-bit recording, aim for peaks somewhere around −18 to −12 dBFS and stop worrying. The old habit of recording as close to 0 dBFS as possible came from 16-bit, where the noise floor was close enough to matter. At 24 bits you have so much headroom that leaving 12 dB unused costs you nothing measurable — and a single unexpected transient that clips costs you the take.
Sampling theory says a sample rate captures frequencies up to half itself. 44.1 kHz reaches 22.05 kHz, comfortably past the top of human hearing. Higher rates — 96 kHz, 192 kHz — give more room above the audible band and push filter artefacts further away, at the cost of file size and processing.
Whether they sound better is chapter twelve. What is not disputed: if you intend to slow audio down substantially, recording at a high rate genuinely helps, because you are lowering ultrasonic content into the audible band. For a sound designer stretching material, that is a real reason rather than a belief.
Bit depth sets dynamic range, not frequency response. 16-bit gives roughly 96 dB; 24-bit roughly 144 dB, which exceeds the range of any practical recording environment. Record in 24-bit. There is no argument here. 32-bit float has become common in field recorders and effectively removes clipping as a concern during capture, which is a genuine practical advance.
Reducing bit depth — 24-bit down to 16 for a CD — truncates values, and truncation produces correlated distortion that sounds worse than noise. Dither adds a tiny amount of random noise before truncating, which converts that distortion into a low-level hiss the ear ignores. Apply it once, at the final reduction, and never twice.
Buffer size trades delay against CPU load. Small buffers give low latency for playing and recording; large buffers let a heavy mix run without dropouts. Use small while tracking, large while mixing. Round-trip latency much beyond about 10 ms starts to interfere with performance, and percussionists notice sooner than most.
Chapter eight
A trainable skill, and the drills that train it.
Engineers do not have better ears. They have better labels. They can hear a problem, name the frequency region, and know what to do — and every part of that is learnable in a way that hearing acuity is not.
Take a familiar piece of music and a parametric EQ. Apply a large boost — 10 or 12 dB — with a moderate bandwidth, sweep it slowly, and listen to what changes. Then have software boost a random band and try to name it. Then narrow the boost. Then reduce it to 6 dB, then 3.
Rough vocabulary to attach as you go: 60–100 Hz weight; 200–400 Hz the mud region; 800 Hz–1 kHz boxiness; 2–4 kHz presence and bite; 6–10 kHz air and sibilance. Twenty minutes a day for a month changes what you can hear more than any purchase.
Sustained listening degrades discrimination measurably. The ear adapts, attention narrows, and — worse — decisions made in that state feel confident. Practical limits: take a real break every 45 to 60 minutes, and treat anything decided after about four hours as provisional. If a mix stops making sense, that is information about you and not about the mix.
The strongest single habit: leave it and listen tomorrow. Almost every mix problem that is invisible at midnight is obvious at nine in the morning.
Chapter nine
Balance, panning, EQ, dynamics, space — and why that order.
The order matters because each stage changes what the next one needs. Reaching for EQ before the balance is right means correcting a problem that would have solved itself.
Dialogue is the priority and everything else negotiates around it. In practice that means carving space in the 1–4 kHz region for voice and letting music and effects yield there, riding levels rather than relying on a compressor to do it, and watching the delivery spec from chapter eleven throughout instead of discovering it at the end.
Chapter ten
What each control actually does to a transient.
A compressor reduces gain when the signal exceeds a threshold. That is all it does. Every other effect attributed to it — punch, glue, sustain, pumping — is a consequence of when it turns down and when it stops.
| Control | What it does | What it does to a drum hit |
|---|---|---|
| Threshold | The level above which reduction begins | Sets how much of the hit is affected at all |
| Ratio | How much reduction is applied above it | 2:1 gentle, 4:1 firm, 10:1+ effectively a limiter |
| Attack | How fast reduction engages | Slow lets the transient through — the stick is preserved and the hit sounds punchier. Fast clamps the transient and the hit sounds softer. |
| Release | How fast it lets go | Fast brings the tail up, adding sustain and room. Too fast and it audibly breathes. |
| Knee | How abruptly the ratio arrives | Soft is transparent; hard is obvious and sometimes wanted |
| Makeup gain | Restores the level lost | The reason compressed material sounds louder |
A slow attack makes something sound more aggressive, not less. Letting the first few milliseconds through untouched and compressing what follows exaggerates the contrast between the transient and the body. Fast attacks on percussion flatten the very thing that makes it percussion. This one fact solves most beginner compression problems.
A compressor responds to the whole signal, so on a full mix the kick drum triggers reduction across everything, and the mix pumps in time with it. Filtering the detector so it responds less to low frequencies fixes this. Nearly every problem described as "the compressor is pumping" is this.
Watch the gain reduction meter, not the output. A few decibels of reduction on individual sources and one or two on a bus is a normal working range. If you are seeing 10 dB and it sounds fine, you have probably stopped hearing it — which is chapter eight's warning arriving on schedule.
Chapter eleven
LUFS, R128, the streaming targets, and the number everyone misuses.
Peak meters measure the highest sample. Ears measure something else entirely. The gap between those two facts produced the loudness war, and a measurement standard ended it.
LUFS — Loudness Units relative to Full Scale — comes from the ITU-R BS.1770 algorithm, which applies a K-weighting filter approximating the ear's response, gates out silence, and integrates over the whole programme to produce one number.
True peak (dBTP) is separate and matters because a signal can exceed 0 dBFS between samples, which a sample peak meter never sees and a lossy encoder can make worse. LRA, loudness range, describes the spread between quiet and loud sections; there is no correct value, but roughly 6–14 LU suggests a master that was shaped rather than flattened.
| Destination | Target | Notes |
|---|---|---|
| EBU R128 (European broadcast) | −23 LUFS, ±0.5 LU file-based | True peak ceiling −1 dBTP. Mandatory by law in some countries; non-compliant programmes are rejected. |
| ATSC A/85 (US broadcast) | −24 LKFS | Enforced via the CALM Act. The one-unit gap from R128 is historical, not perceptual. |
| Netflix | −27 LKFS, dialogue-gated | Different gating, so not directly comparable |
| Spotify, YouTube, TIDAL, Amazon | ≈ −14 LUFS | Playback normalisation |
| Apple Music | ≈ −16 LUFS | Widely reported Sound Check measurement, not a figure Apple publishes |
| Podcasts | ≈ −16 LUFS |
−14 LUFS is a playback reference, not a production target. Platforms measure your file and apply a gain offset at playback; your file is not altered. Mastering to −14 to "match Spotify" is unnecessary, and crushing dynamics to get there is actively harmful, because a louder master is simply turned down to the same place with its dynamics already gone. Some platforms turn quiet material up and some do not, so a very quiet master can sit low against everything around it — but that argues for sensible loudness, not for maximisation.
Broadcast is a different matter entirely: −23 LUFS is a specification, and a programme that misses it gets rejected.
If you deliver to European broadcast, measure integrated loudness against −23, keep true peak at or under −1 dBTP, and check with a compliant meter rather than trusting a plugin's rough reading. If you deliver to streaming, make the mix sound right and let normalisation do its job.
Chapter twelve
What is audible, what is not, and what nobody has actually tested.
This chapter will annoy people in both directions, which I take as evidence it is roughly in the right place. Three categories, kept separate on purpose.
This deserves care because both camps misreport it.
Meyer and Moran's well-known 2007 study in the Journal of the Audio Engineering Society inserted a 16-bit/44.1 kHz conversion loop into high-resolution playback and found listeners could not reliably detect it. That result was widely taken as settling the matter.
It did not. A 2016 meta-analysis by Reiss, also in the JAES, pooled a large number of trials across many studies and found a small but statistically significant ability to discriminate high-resolution from standard-resolution audio — an effect that improved with training.
The honest summary: the effect, if real, is small, requires trained listeners, and is nowhere near the transformation the format's marketing claims. It is also not zero, and people who declare the question closed are overstating what the evidence supports. Both of those sentences are true at once.
A great deal of audiophile practice has simply never been through a controlled listening test. Not disproven — unexamined. Component isolation, mains conditioning, various tweaks: the studies mostly do not exist, and the reason is that proper blind testing is slow, expensive and unglamorous, and nobody has an incentive to fund it.
Treat "no evidence" and "evidence of absence" as different statements, because they are.
Apparatus
Standards, papers, and the blind tests.
House rule
Stated as plainly as what I could.