The ultimate loudspeaker-in-room listening test — reflections and directivity v1.58

Walls & ceiling — a low / high

Two absorption coefficients per surface, low band and high band. The split is the same one the reflection shelf and the two decay times of the tail use, so one number cannot drift from the other. Typical published values for that kind of construction rather than measurements of any particular wall – the same standing as every other number in this model.

The order of the two numbers matters more than their size.

• Bare concrete absorbs almost nothing at either end: the tail stays bright as it dies.
• Curtains and carpet absorb the top far harder than the bottom: the tail goes dark and the bass outlasts it.
• Panelled surfaces run the other way round. Gypsum board and plywood on studs resonate and absorb the low end harder than the high one, which is why a stud-wall living room sounds brighter and thinner than a masonry one of the same size, and why it needs no bass trap so much as it is one.

The low figure is the published 125 Hz value, not an average across the band, and that is deliberate. Its most audible consumer is the modal region below the Schroeder frequency, where the mode bank gets its decay and its Q – and a stud wall really does absorb that much down there, which is why such a room has less bass boom than a concrete box.

The cost: the same figure colours reflections and the low half of the tail all the way up to the split, where 125 Hz is the wrong reference. A gypsum panel is far less absorbent by 1 kHz, so its reflections come out about 1 dB darker than published data would give; a carpet or a curtain is far more absorbent, so theirs come out about 2.5 dB brighter. That is the gap between the step and the curve in the charts below. Fixing it would take a third band.

Both numbers are drawn below against published octave band data for the same kind of surface. The line with dots is the published data, 125 Hz to 4 kHz; the flat two-step line is what the model uses, one value below the split and one above. Where the step sits away from the curve, the curve is the published figure and the step is the simplification. The two surfaces you are listening to right now are marked in use.

Handbooks disagree with each other by a good deal – the two standard rows for heavy carpet differ by 0.18 at 250 Hz on their own. Two rows have no standard table entry at all and say so. Read them as a sanity check, not as a datasheet.

Late tail — Reflections n % of the room budget · the tail fills the rest

100 % on the slider is not the loud end of a preference. It is the level that brings the whole room side – these reflections plus the tail – up to what a diffuse field of this room's absorption would be. The slider is a percentage of that calibrated amount, and the line says how much of it the early reflections already supply.

• The early reflections are subtracted from the budget, not stacked on it: the tail gets √(budget − early), the early energy averaged over listening positions – only a room average may be taken off a room average.
• The split moves with the speaker, which is why it is printed beside the slider: a more directive speaker puts less power into the room, so it both leaves the tail less of the budget and makes the tail quieter. In the same room the reflections read anywhere from well under a quarter of the budget to most of it on that choice alone. Absorption lowers the tail as well.
• The channels are decorrelated, so two speakers give √2 of the tail.
• A bigger room gets a quieter tail, and it is the budget itself that shrinks rather than the sharing: the critical distance grows with the room, and the budget is the diffuse level at that distance. The reflections take a slightly larger share of that smaller amount, not a smaller one.
• Resizing does not stretch the tail you are hearing: a new one is built and takes over once the dimension settles, so the change arrives as a single step rather than following the slider.
• Capped on the line means the room is live enough that the honest level would clip the output, so you are reading the room as quieter than it would really be – pick a more absorbent wall or floor material.
• Held at a floor means the reflections alone already reach the diffuse estimate.

Turn — Left / Right / Face speakers

Hold a button to turn at a rate scaled from Move speed. The angle wraps, and Head follows greys all three out.

Face away and listen for whether the image follows you. Front-back confusion is a known failure of a generic HRTF, and hearing it is what this control is for.

Lock — Speakers follow · Head follows

• Speakers follow – your move is applied to the speakers as the same delta, so the triangle keeps its shape and the whole set-up travels through the room: where you sit becomes separable from how the triangle is shaped. Centre greys out, since it would carry the triangle with it and centre nothing.
• Head follows – the direction is recomputed from the speakers on every update, i.e. Face speakers held down rather than pressed once, so walking never turns into turning. Target is the pair, or the one speaker still playing if you mute the other: muting as an A/B therefore compares one speaker on axis against two in front of you, not the same seat twice. The Head slider and the Turn buttons grey out.

Neither lock stores a value of its own – the sliders keep showing the real position, and switching off leaves everything where it is. Height, Toe-in and Ear height are outside both.

Speaker library

A polar export becomes a named entry in the list. Files are read in your browser; a measurement reaches the server only if you back it up to your account or tick it into a shared link. It plays as soon as it is imported. File format →

• Required: the on-axis file hor 0. Everything else is read against it.
• Half a circle is enough horizontally – the model reads the angle’s magnitude. Uneven angle steps are fine.
• ver files are optional – without them the vertical polar is copied from the horizontal one.
• The line under the file picker reports what arrived: angles per plane, range, step, widest gap.
• Export at your listening distance. The examples here are 3 m, which stops being honest much below 2 m. Your ears have to stay level with the speaker – the height lock keeps that true.
• A speaker arriving with a shared link or a preset file plays the same way. With an account, Back up copies one to your account so another machine gets it.

Description travels with the speaker: export files, the copy in your account, shared links. It is what tells someone opening your link what they are hearing, since the name does not travel. Private memo is backed up to your account only – never exported, never shared.

The spheres have no design choices: a point source on a sphere, size the only variable. Their pattern depends on ka = 2πr/λ, so the whole set is one curve slid along the frequency axis – double the diameter and the baffle step moves down an octave.

The classic three-way is the only asymmetric entry: mid and tweeter sit off the centre line of a wide baffle, so it radiates differently towards each side wall. The data is the left cabinet; the right plays it mirrored. The horizontal map draws both halves instead of mirroring one, and the note under it names which case is loaded.

Polar file format — plane · angle · three columns

Plain text. VituixCAD writes it directly (File → Export → Polar Frequency Responses), but nothing about the import is specific to it: a measurement rig, a BEM/FEM solver or a script of your own work the same way once the names and columns match.

File name: only the ending matters – hor or ver, the angle in degrees, .txt. Anything may come before the plane; between plane and angle anything but a digit or a minus sign. So my-speaker hor 0.txt, sim_ver_-30.txt and VituixCAD_PolarFR hor 45.txt all read. One file per angle, and you pick the whole set at once.

Contents: one line per frequency point. Hertz, then dB, then phase in degrees if you have it. Separator: space, tab, comma or semicolon; decimal point or comma. Any line that does not start with two readable numbers is skipped, so headers and blank lines can stay.

# my-speaker hor 30.txt — a line that is not two numbers is skipped
# Freq(Hz)  SPL(dB)  Phase(deg)
20.0      -8.41     -12.3
25.0      -8.36     -15.8
31.5      -8.30     -19.9
...
20000.0  -31.07     146.8

• Frequency spacing is yours. The page resamples onto its own band grid, taking the energy average within each band, so a dense set loses nothing and a sparse one is interpolated across.
• hor 0 is required. Every other angle is read against it.
• Phase is all or nothing per file. One unreadable third column and that file is magnitude only. Wrapping at ±180° is fine – it is unwrapped for you.
• Beyond the angles you export, the edge value is held rather than extrapolated – the picker line names it when the coverage falls short of the steepest reflections.

What the third column buys: each angle sums with the right timing, so the reflections and the two speakers add as the real thing does off axis. Without it the directional phase falls back to minimum phase – not to the on-axis phase – which is right for a formula-built source but not for a real cabinet; the difference shows in the crossover region and towards the sides. On-axis phase and per-angle phase are separate: a third column in hor 0 alone gives the first, the whole set gives the second.

Synthetic radiator — n° · holds to n Hz

A source built from a formula, not measured: you set a −6 dB total angle and a physical size for each plane. It is convolved exactly like a measurement, so the graphs read the data the sound uses. It answers a building decision: how narrow does the vertical pattern have to be before the floor and ceiling stop mattering, and is an object that size already too big to build?

The number on each line is where that plane’s pattern stops holding. Below it the object is small compared with the wavelength and the pattern widens towards a sphere. A wide pattern needs a smaller object than a narrow one – hence two size sliders, and the vertical number is usually the awkward one.

What it is not:

• The axis is flat by definition – the sliders change directivity and nothing else. There is no cabinet response here.
• An ideal aperture: no diffraction around it, no cabinet behind it, and its directivity levels off high up where a solid object of the same size would keep narrowing.
• No rear lobe. A real horn or sphere radiates a peak straight backwards; this does not.
• Each pattern holds only as far down as its own dimension allows – which is the point, not a shortcoming.

It will not sound like a speaker from the list even at the same size, and the reason is frequency rather than size. A real cabinet’s pattern keeps moving: narrowing as the driver grows large against the wavelength, widening across a crossover where a smaller driver takes over, narrowing again above it. This one settles on the angle you set and holds it. A room hears a speaker mostly through its reflections, so a pattern that stays put and one that moves colour it differently – audible long before either is wrong. Sliders to hear what a chosen pattern does to a room; the list when you want a cabinet.

Response — Directivity only and three phase models

A polar export holds two things. Directivity is every angle stored as its difference from straight ahead – which is why 0° is exactly flat. The on-axis response is the part that was divided out. First position: directivity only. The other three add the response back, each with a different phase.

• Minimum phase – what the magnitude alone dictates; the phase any smooth passive filter of that shape would have.
• Measured phase – adds what the export measured: the part a crossover adds in time and which does not show in the magnitude. A Linkwitz–Riley pair sums to an all-pass, so equalising cannot find it – it has to be read from the phase column.
• Linear phase – removes both, on the axis. A passive crossover cannot do it, an FIR can; the switch is that comparison without the build. On the axis it rings before the transient, which is what linear-phase crossovers do and why the position exists. The other three do not.

Shape, not level: the average energy between 200 Hz and 10 kHz is removed, so the response changes tone and not loudness, and all four positions share that curve exactly. A constant delay is fitted out of the phase over the same band, so the list changes dispersion and not arrival time — the measuring distance does not come with it. Latency is the same in all four, so switching moves nothing in time. It is applied once per speaker, ahead of the room, so the walls and the modes reflect it. Bypass under the spectrum does not touch it – that button compares your curve.

Phase is read at every measured angle, not only on the axis, and it does not stay the same off axis. On the ideal two-way, sideways it stays within 0.07 ms of the axis out to 60°; vertically it changes sign across the crossover — at 30°, +0.04 ms below the woofer’s band and −0.10 ms above it, because the drivers are not on the same axis. The direct sound leaves along the axis; the floor and ceiling bounces leave at 32° and 48° in the default room and carry the phase measured there. How much it matters follows driver spacing – least on an ideal two-way, most on the classic three-way.

• Linear phase is linear on the axis alone. One filter at the input, ahead of the drivers and the room, straightens the direction it was measured on; every other angle keeps the difference it had, including the bounce angles the correction never saw. Not a speaker that is linear everywhere, but the one you could build, heard from where you sit.
• Between measured angles the page interpolates, so a 10° export is a 10° model.
• No phase column, none of this: each angle’s phase is built from its own magnitude.
• Under Shorter filters the reflection filters cannot carry phase the magnitude does not dictate, so they stay minimum-phase whatever you pick.
• Two different crossovers on the same speaker is a different question, and this list does not answer it alone. The difference shows as vertical lobing, magnitude and phase both – the microphone saw the drivers summed at each angle, so what is missing is not a source per driver but whatever the export left out. Two exports of the same speaker give different power responses and different reflections. LR2 against LR4 is that comparison as far as it goes today.

A position is grey when the speaker cannot do it, and the line underneath says which case: an ideal pattern has no measured response; a measurement imported before this existed kept only the magnitude, so it gets minimum phase until you import its polar files again. That line also carries the number that decides whether any of this is worth hearing: the largest excess group delay between 100 Hz and 1 kHz, after the constant delay is out, so it counts dispersion and not distance. The example two-ways read a fraction of a millisecond; the three-ways read two orders of magnitude more and low down, where a steep high-pass and two crossovers put it.

The critical distance estimate does not follow the switch – it assumes a flat speaker, and the line below it says so whenever the response or the EQ is shaping anything.

Spectrum — an omnidirectional microphone at your seat

The direct sound, the reflections and their comb filtering, the late tail and the room modes, in one curve at your listening position.

• The tilt of the curve belongs to the source, not the room. What the room does is the departure from it – watch a null appear as you walk, and move when you change the room.
• Read before the binaural stage, and summed to mono there. Two ear signals that arrive a fraction of a millisecond apart sum to a comb filter neither ear hears, so it would put ripple in the top octaves of an anechoic direct sound. The curve is therefore the pressure at your seat rather than at an eardrum.
• The headphone correction is not in it. That sits after the master output; this curve is the loudspeaker in the room.

Headphone correction — Load · Paste · Bypass

Last in the chain, after the master output, and not part of what is modelled – it corrects your headphones, not the room. The room meters and the spectrum curve do not see it, for the same reason.

The file is AutoEq’s ParametricEQ (AutoEq publishes one for most headphones). Load a file or Paste the block straight from that page – same parser.

• It stays in this browser. Not in an export, a room saved to your account or a shared link, and opening someone else’s link does not replace yours. It survives a reload, and the row names the file.
• Reset does not clear it – that resets the model, not what you listen through. Clear on this card removes it.
• Reads PK, LSC and HSC rows. Other filter types, a missing preamp line and a GraphicEQ file are named on the row rather than dropped in silence.
• Web Audio’s shelving filters have a fixed slope, so a shelf Q other than the standard one is not played as written. AutoEq writes the standard one; a hand-tuned file may not, and the row says so.
• The curve is drawn from the filters themselves, preamp excluded – that is what makes it comparable with AutoEq’s own graph.

Bypass keeps the correction’s preamp by default, and that is a measurement decision. A correction is built with headroom below it; without that preamp, switching it off would also raise the level – and louder sounds better, so the comparison would answer “which is louder” when you meant to hear the shape. If the bypass still sounds louder or quieter, the trim beside it settles it.

Speaker EQ — Bypass (e)

The curve is drawn over the spectrum, and you edit it there. Drag a dot sideways to move the band, up and down for its gain, scroll on it for Q. The knobs below the graph are the same numbers, so either one follows the other. The dots marked HP and LP move sideways only – a pass filter has no gain to drag. Touching any control switches the EQ on; the default is bypass.

The knobs turn by dragging up and down, not in a circle, and shift makes the travel finer. The wheel steps a fraction of the range over a knob, or a single step of the control with shift. The unit is in the label above each knob, so the reading below it is a bare number: 20.0k under Hz means 20 kHz.

Double-click any slider or knob on this page to put it back to its default. Nothing snaps to zero, deliberately – snapping would make the values either side of it hard to pick, and that is where fine adjustment matters most.

The graph carries two vertical scales at once. The orange spectrum is a level in dBFS, labelled on the left; the blue EQ curve is a gain, labelled on the right. They share the frequency axis and nothing else, so the EQ curve sitting above a peak does not mean it is louder than it – it means it is adding that many decibels there.

The EQ sits ahead of the speaker and the room, so the walls, the floor and the room modes all reflect the curve you set – boost the bass and you hear what the mode does with it, not just what your ear does. That is the reason for an equaliser on this page at all. Both speakers get the same curve; this is a listening test, not a channel-by-channel calibration.

Bypass is on the E key, so you can compare without the mouse. It is a real bypass: the signal goes round the filters rather than through flattened ones. Flat resets the curve and leaves the bypass where it is.

• Parametric bands, plus a high-pass and a low-pass whose Q is in decibels – that is Web Audio’s unit for those two filter types, and Butterworth is −3.0 dB rather than the 0.707 you may expect. A band with 0 dB of gain is transparent, so there is nothing to switch off.
• Slope 12 or 24 dB/octave, and the high-pass is named after the enclosure: a sealed box rolls off at 12, a reflex box at 24. Set it to your box’s corner and you are listening to that alignment in this room. There is no "reflex" low-pass, so that one is labelled by the number alone. Either slope is −3.0 dB at its own corner frequency.
• Both reach well past the bass, so the pair is also a band-pass: hand the room a single octave and hear which reflections belong to it, or listen to a driver’s own passband. Drive them past each other and nothing comes through — the page says so rather than leaving you to wonder.
• Preamp is headroom, not tone. Boosts push the output towards clipping, so Peak boost sets the preamp to the tallest boost of the curve. The row reads the Peak, the preamp and the net left after it, and warns when that net is high enough to clip. Sustained clipping sounds like a distorting driver, which is exactly the thing a test like this must not invent.
• One reading does not follow the EQ: the critical distance estimate knows only absorption and directivity, so it still assumes a flat speaker – and says so on the line below it while the EQ is shaping anything. Measured direct vs room does follow it, because that one is measured downstream of the filters.

Ear height — At speaker height

The browser stores its head-related filters on a 15° grid of elevations and truncates to the one below the source, so a boundary sits at 0° and every 15° from there. Crossing one changes the tone in a single jump and leaves it changed: −6 dB at 6.4 kHz for two millimetres of ear height, measured here. That boundary is where a listener normally sits, level with the speaker, and a reflection in a vertical wall keeps the source's height – so the direct sound and every wall reflection cross it at the same millimetre.

Ticked, your ears stay level with the speakers: elevation is zero and nothing is ever crossed. Unticked, every path runs through two filters from neighbouring bins, crossfaded by where the real elevation falls between them. That takes the step down to the noise floor of the measurement, and both positions are rewritten for every path each time you move – a load switch rather than an accuracy one: if the audio stutters, tick it back on.

Two limits it does not remove. Floor and ceiling reflections cross their own boundaries as you walk, because their elevation depends on distance. And ticked it is a constraint: a vertical listening test is not done with it on.

Vertical + — + angle points up

Only appears for a measured speaker: the export does not answer it, and the synthetic radiator is symmetric vertically, so there the setting would do nothing. A measured vertical set usually is asymmetric – a woofer below a tweeter does not radiate the same way up as down – so which sign means up decides whether the floor reflection or the ceiling one gets the off-axis response that belongs to it. Backwards swaps the two. Up is the owner's convention here; the file itself only says +30°.

It does not move Q on the critical distance row: that integral reads both halves, so the sign only exchanges them.

Directivity check — the measurement against what plays

The measurement is the thick line, what you hear is the dashed one. They are the same data: each path convolves the measured curve directly, so the dashed line is the measurement at the nearest angle cell rather than a model of it. Where the two separate, that separation is the rounding.

The table is that separation as a number – dB rms over every measured band, for the speaker playing right now: whole circle, front, below 250 Hz. Near zero on axis, because the pattern barely moves there; largest where it turns fastest, i.e. the deep off-axis nulls above 1 kHz and the rear. Worst may be the filter’s floor instead of the rounding: where the measurement has a deep null the chain stops following it rather than approximating it.

The Response switch changes what the curves are. The polar data is normalised to the on-axis response, so on Directivity only 0 dB is the speaker’s own axis at every frequency – directivity alone, plus your EQ, which stays in every view because it sits ahead of the speaker. A speaker aimed straight at the seat therefore draws a flat direct line, or the shape of your EQ: that is the zero level looking at itself rather than a perfect loudspeaker. Switch the response on and the on-axis curve is added back into At your seat, All angles and the power response – 0° is no longer flat, and the dips are the ones the chain actually produces. The line under the graph says which you are reading. Both maps stay directivity: the error map is the measurement minus the cell that plays, so adding the same curve to both would not change it.

• At your seat – the heavier pair adds the room: six early reflections summed with the direct sound as complex pressure, delays and all, read against the direct sound’s own level. Smoothed, so the finest detail is missing; below the crossover the room is carried by the mode bank, which is not in this graph; and it is the pressure at the seat, not at your ears, where the head and the HRTF would come in.
• All angles – the white pair is the power response, the energy average over the whole sphere, which is what a room ends up hearing. Same line in both planes, because the integral uses both. Measured data hides the measurement so you can read the playing curve alone.
• Towards each surface – what leaves the speaker towards each wall, the floor and the ceiling, not what reaches your ears. That is where placement is decided: an off-axis dip aimed at the wall beside you arrives as a reflection with a hole in it, and a few degrees of toe-in moves the hole. Mute that surface and hear whether it matters. Left speaker only – the right is the same curve mirrored unless you have moved something, and the asymmetry is what At your seat is for. Wall absorption is left out: it is smooth and would not put a dip anywhere, so it would only leave you wondering whether a hole belongs to the speaker or to the carpet. Each curve follows its own surface switch above, and with the response on the white direct line is the real thing and not flat.
• Polar map – a map of directivity, named as one.
• Error map – the measurement at the exact angle minus the filter you get, which is the nearest cell played through a short filter built on the phase model you chose. Flat grey means it cost nothing there. Most is rounding, but the map reads the real filter, so it also shows that filter’s two limits: its length, and a floor below which it cannot follow the measurement. The chequered patch in the rear top octave is that floor, not rounding, and it is where the table’s largest number comes from. It has its own key beside the colour scale rather than a stronger colour, because it is a different thing rather than more of the same one. The note under the map names the cell, the filter and the band density, and says which of the two you are looking at.

An ideal pattern — omni, cardioid, dipole — is a plain gain per angle rather than a measurement convolved into the signal, so there is nothing to draw beside it.

Reflection detail — Full · Shorter filters

Fourteen paths each convolve a directivity filter. Shorter filters cuts the twelve reflection paths to a small fraction of their length — the line under the switch says both lengths — and leaves the two direct paths untouched, because the direct sound is the one you localise on. What you give up is the directivity curve itself, at every frequency and not just at the bottom. Use it when the audio crackles.

It also pins the reflections to minimum phase. Phase that the magnitude does not dictate needs room ahead of the peak, and the shortened filter has none, so with this switch on the phase position you pick applies to the direct sound alone.

What it costs depends on the cabinet, not on the setting. A filter length is a length of time, and a wide baffle puts its directivity structure low, where structure is long. Measured back when the full length was shorter than it is today: 0.17 dB rms on the bookshelves, 1.8 dB rms on the 1 m sphere, the spheres passing the bookshelves at about 38 cm. Against today’s full length the gap is wider, and it has not been measured again. Judge it on the speaker you are listening to, not on the list.

It does not fix the stutter while you walk. That is not convolving, it is building: every time your angle rounds into a new cell the page builds a new filter, at full length either way. If walking stutters, the levers are the room and the paths – mute surfaces you are not judging, turn Modes off, or move with the sliders instead of holding a key.

Where you can see it. The polar map and the error map always read the direct path, so they do not move when you switch. The curves under Towards each surface are drawn from the shortened filters, so those do – that is the window into what the setting costs.

Both figures are measured across the whole speaker list, both planes and both sides of the horizontal where a speaker is not left–right symmetric, at five filter lengths. Full length is the default because of the large cabinets.

Group delay — the crossover against the room

The phase of the Response switch drawn as time. The white curve is the group delay you hear on measured phase, the grey curve the one on minimum phase, and linear phase is the zero line itself — every frequency arriving together on the axis. The gap between white and grey is the whole of what the switch does: if they lie on top of each other, that A/B has nothing to find on this speaker. It needs a measured speaker whose export carried the phase column, and says so when it has none.

The dashed blue lines are this room. Each one is how much later a reflection reaches you than the direct sound, from the same geometry the sound uses — move a wall or walk, and they move; mute a surface and its line stays, greyed and labelled. They are there because milliseconds on their own do not say whether they are many: a crossover that delays the bass 2 ms is a different proposition in a room whose nearest reflection is 1.5 ms behind than in one where it is 15.

⚠️ They are not the same kind of millisecond, and the graph is not claiming they are. A reflection is a second arrival of the whole signal; group delay is a smear inside one arrival. Equal numbers do not mean equally audible — the comparison is there for the order of magnitude. Whether either is audible is what the listening test is for, and the square-wave source is the one that makes the smear easiest to hear.

• Two scales. Room scale fits the nearest reflection into the picture, which is the comparison. Speaker scale drops it and fits the curves, which you need when the answer is “the speaker does almost nothing” and the curve would otherwise be a flat line along the bottom. The line under the graph carries the numbers either way: the excess delay across the midrange, the nearest reflection, and any reflection above the top of the scale.
• On axis only — the stored phase is the on-axis phase, so there is no horizontal or vertical choice here and those buttons are hidden. Off axis the page uses the same phase, and how good an approximation that is measured is in the Response section.
• The curve stops short of the ends of the measurement, because the window it is read over has to fit inside the measured band. Beyond those ends the model holds the last value it was given, and a group delay read off that would describe the model rather than any loudspeaker.

Reading the pair: as far as the grey curve follows the white one, that much of the delay is the magnitude’s own and would be there in any speaker rolling off that way; the rest is what only measured phase has. Where the white curve crosses a reflection line, the room above that frequency is doing more than the crossover.

Ears — Binaural (HRTF)

On, the browser's own head-related transfer function does the listening: interaural delay and ear filtering, which is what makes height and front-back audible on headphones. Off is plain amplitude panning – direction still follows where you face, but nothing goes through a head. Measured with the head at 90°, switching off takes the interaural delay from 542 µs to zero and puts the first arrival back on plain geometry.

Two things it is for:

• Separating the HRTF from the room. Off, the floor and ceiling reflections are level and delay only, and the elevation grid does not exist, so height comparisons run continuously (ear height →).
• Rendering an impulse response for loudspeakers. The same switch decides what the file below records, so the file matches what you listened through – and a binaural IR played over speakers passes through a head twice.

Expect a tone shift, not only a change of placement – and it depends on direction. Both ears summed, referenced to 500 Hz, against plain panning: a ±30° pair reads +3 to +4 dB below 250 Hz, within a dB from 2 to 5 kHz, then −5 to −10 dB from 6 to 10 kHz; over the whole sphere the 2–5 kHz range reads −4 to −5 dB instead, so sideways reflections are coloured differently from the pair in front of you. Binaural sounds thicker and less airy for that reason – the head and outer ear doing their job, not a fault, though on headphones it arrives twice since target curves already contain an average outer ear. And because the colouration carries the direction, no fixed compensation curve removes it without making some directions wrong.

Critical distance — n m · Q n · R n m² · seat n dB

The distance at which a real room with this absorption would have its reverberant field overtake the direct sound: r_c = √(Q·R/16π), with the room constant R = S·ā/(1−ā). It is theory on this room's absorption, not a reading from the mix you are hearing – the late tail here is a slider and the reflections are first order only, so the model's own ratio at your seat is a different number. That is the seat figure. Compare the distance with your own room rather than with what you hear here.

It does not depend on how many speakers play: a second one adds direct and reverberant power in the same proportion. The measured ratio does, by about 3 dB, because two mono speakers sum coherently on the centre line while their reverberant fields do not.

The line prints R rather than the more familiar RT60 on purpose. RT60 is clamped for the tail, so at extreme absorption the printed RT60 and this distance would no longer describe the same room. R is exactly the number r_c is computed from, so the arithmetic checks from the line itself. The rule of thumb 0.057·√(V/RT60) is the same formula without the (1−ā) term, and it reads low in a furnished room and further off in a dead one.

Q is closed form where it is closed form: omni 1, cardioid and dipole exactly 3 – both have a power integral of 4π/3, so they differ in where the energy goes, not in how much of it goes off axis. A measured speaker is not one number, hence the extremes band by band on the second line. A value below 1 in the bass is not a fault: normalised polar data is slightly above 0 dB off axis down there, so the cabinet radiates a little more to the sides than on axis.

Measured direct vs room — n dB · n dB wetter

Read from the running audio, not calculated: two analysers sit in the graph, one on the direct sound and one on the room side (every first order reflection plus the late tail). The number is the ratio of their powers as a running average, so it lags while you drag, and it only exists while something is playing – hence press play when it is not.

It is taken above the crossover and without the mode bank, because below the crossover there is no separate direct sound to compare against, and before the HRTF, because the interaural delay would comb filter the sum and the meter would show a dip neither ear hears.

The second line compares it with the seat estimate on the critical distance row, and a residual is expected rather than an error. The late tail has the average early energy subtracted from it – averaged over listening positions – so a seat closer to the speakers than that average still reads wet, usually within a decibel: near them the early field really is stronger than a room mean. More than a decibel or two of disagreement with the tail at 100 % is the geometry talking – a seat that close to a wall is not a diffuse field. It is not an invitation to pull Late tail down. Reading drier means the opposite: the tail is below its calibrated 100 %, or surfaces are muted.

Three short words can appear instead of a number:

• silent – nothing is coming through above the crossover.
• direct only – the room side is muted, which is the anechoic reference.
• room only – the direct sound is muted.

Stereo triangle — n°

The angle at the listening position – the one people mean by the well known "60 degree triangle", where each speaker sits 30° off your nose. Toe-in is a different angle: how far the speaker itself is turned towards you. The two are independent, so an equilateral triangle with the speakers pointing straight down the room is a setting you can have. Aim at listener names the toe-in that puts the axis on your ears.

Distance to the speakers — Speakers n / n m · Δ n mm (n ms)

Equal distances mean you are on the centre line, and keeping that is the whole reason the walk buttons exist: they change distance only, so the lateral position is preserved exactly. A mouse drag drifts sideways, and the interaural time difference drifts with it – the image smears while you judge it.

The delta is in millimetres and in milliseconds because the second one is what you hear: a few centimetres is a fraction of a millisecond, already an image shift rather than a level difference. It is also why the walk is slow – the delay lines retune as you move, and at the default speed that pitch shift is a fraction of a cent; at metres per second it would be audible.

A / B — Now A · A and B differ in toe-in, separation

Store A and Store B put the current settings in two slots; A / B – the button or the space bar – switches between them. The line under the buttons names which one you are listening to and what the two differ in. An empty list means the switch is silent, because there is nothing to hear.

What you are listening in is shared; what you are listening to is not. Once A and B agree on a setting of the first kind, adjusting it is written into both slots as well as into what you hear – nothing is thrown away for a change you never meant as a comparison.

• Shared: the room and its materials, which of its surfaces are live, the room modes, where the speakers stand, where you sit, which way you are facing, the movement locks, the walking speed, muting one speaker of the pair and cutting the direct sound.
• Compared: the speaker itself, its directivity, its own response and phase model, the equaliser and the binaural mode. These stay in the slot you stored them in even when both slots agree, because a shared setting cannot be compared – choosing a different speaker would change both slots at once and leave nothing to switch between. A change to one of them counts as an unsaved edit, the line says so, and storing it in the other slot is what turns it into a comparison.
• In neither: the listening level, the source, the sine frequency and the headphone correction. If the test signal changed with the switch you would be hearing two differences at once.

The number keys are more of the same, for when two is not enough. Press one to recall a setup, hold Shift and press it to store the current one there, and 0 to flip back to the one you had before – a shelf of loudspeakers compared one after another is what they are for. The line under the buttons lists the ones that have something in them. They hold what A and B hold, but they live in this browser instead of in the page: they survive a reload, and they do not travel in a shared link. A and B are what a link carries.

Paths — RT60 n s (LF n / HF n) · mean free path n m · n paths

RT60 is Sabine on the current room and materials. The LF and HF pair is what the synthetic tail decays with – the broadband figure is the reading, and the length of an exported impulse response. Mean free path is the average distance between bounces, and it sets the pre-delay of the tail. The count is the direct sound plus one first order reflection per surface, per speaker. Mute a surface and the path stays: muting removes the energy, not the path.

Room modes — Modes n · Schroeder n Hz · Q n–n · axial n / n / n Hz

Below the Schroeder frequency the image source model is replaced by an analytic mode bank – one bounce per surface cannot produce a mode, and a mode is what is left after infinitely many. Each mode is a second-order low-pass whose level is its pressure at the speaker multiplied by its pressure at your ears, so a zero means you are sitting in a null. The filter type is what makes the nulls possible: phase runs from 0 to −180° through resonance, so two adjacent modes on opposite sides of their own resonances are out of phase and cancel. A bandpass never changes sign, and modes could only add. Q comes from the low band RT60, the LF figure on the row above.

Q also sets how high each mode peaks, not just how narrow it is: a term of the modal expansion peaks in proportion to Q/f ², and Q itself grows with frequency, so the two leave a net 1/f tilt. That is the low end a room adds and a field does not, and why taking a speaker outdoors takes the bottom with it.

The term with all three indices zero is not a mode: every cosine equals one, so it does not vary with position. It is the room's pressure below its first mode, and it is what you hear when you switch the room off and the bottom octave changes. Low-passed at the lowest mode above, flattened at the bottom of the model's range below – a real room leaks, so the pressure rise cannot continue to DC.

The axial trio is the lowest width, length and height mode, c/2 divided by each dimension: the ones you can walk into. Capped on the line means the bank ran out rather than the room – the modes above that frequency are missing from what you hear.

A symmetric pair fed the same signal cannot excite a left–right antisymmetric mode, and that is a little under half of the bank. While Off centre is at zero, every odd width index – the lowest width mode included – gets equal and opposite force from the two speakers. Toe-in does not change it and neither does where you sit; the mode is silent from every seat. Sliding the pair off centre does, and so does a recording whose channels carry different bass. It is also why real rooms do not sound like this: real placement is never quite symmetric and real recordings are not mono.

Settings — Export JSON · Import JSON · Save to server

All three carry the same object out of the browser – the room, the seat, the speaker choice, the surfaces, the EQ and your A and B snapshots. Import reads one back.

• Export writes a file you keep. A measurement you imported yourself travels in it, with its description, so whoever opens the file hears the same speaker without owning the measurement. One of this site’s own speakers travels as a name instead — the reader’s page has it already.
• Save to server keeps a named room on your account, where it opens on your other machine and survives a cleared browser. It needs a free account and an email address; exporting and importing need neither. How many rooms an account holds is shown under the list rather than promised here, because the server is what decides it.
• What stays behind, in all three: the listening level, the source, the sine frequency and the headphone correction. A speaker’s private memo is backed up to your account only – never exported, never shared.
• Nothing from this page reaches the server until you save a room, share a link or back up a speaker.

Share a link

A shared link carries the settings, not audio: the room, the seat, the speaker choice, the surfaces, the EQ and your A and B snapshots. The reader’s own browser renders it, so what they hear is what you heard, computed again on their machine. Their listening level, source and headphone correction stay theirs — a link does not carry those.

• Reading needs no account. Making a link needs a free account – an email address and nothing else – and that is what gives the link an owner: it stays yours, and deleting your account deletes every link you made with it. Taking a link down stops it working everywhere you have pasted it, including that forum thread.
• Your own speaker measurement travels only if you tick the box that says so. Without it, a link that used one of this site’s speakers still works – the reader’s page fetches the same speaker by name.
• A link nobody opens expires; one that keeps being opened does not. The clock runs from the last opening. You see how many times a link has been opened, and that is all there is: a count, no addresses, no per-visitor record.
• Pressing the button twice does not make two links. The link is made from the settings, so as long as those have not changed you get the same address back, with the message updated if you edited it meanwhile. Change the room and you get a new link, because it is a different room. Typing a second address of your own is the way to ask for a second link to the same settings.

Pro names it. A Pro link has an address you choose (/preset/your-own-address), a short message shown on the page, your nickname and one link back to your own thread or site. The message is the useful part: “listen to the 45° ceiling reflection and tell me if you hear it” is a reason to open a link, a pile of slider values is not. The link back is rel="ugc nofollow" – it is there for people who click it, not for search engines, and this is said out loud so that nobody buys Pro for the wrong reason. Save text edits those afterwards without the address changing, because by then it is already in somebody’s post. Pro links do not expire, and they keep working after a Pro period ends: a dead link would punish the reader, not the person who stopped paying.