Research: which dips EQ can fix, and what the body (not the install) is responsible for — pooling cabins

[ayukhno]

This one is not a feature request and needs nothing from the code. It is the research the tooling was built for, where it is stuck, and an offer to pool what we each have — because the wall is the same for both of us: field data from more than one or two cars. You wrote that most of your field-derived heuristics come from three cabins; ours come from one. That is the honest state of the art in this corner of audio, and it is why a lot of what tuners "know" is folklore.

1. Which dips an equaliser can fix — and which it must not touch

Every EQ decision in a car comes down to one question: is this dip a shape deficit (fixable) or a cancellation (an interference null that EQ only makes worse, burning headroom and excursion for nothing). Your EqBoostabilityMask answers it with coherence and narrow-deep-null detection; ours (rew_tool/eq_gate.py) answers it with the excess phase: REW's own excess-phase decomposition, a sliding RMS of the excess-group-delay z-score against a ±1-octave baseline, ANDed with local depth and with the candidate filter actually delivering gain there.

Where it stands, honestly:

• on an analytic family of minimum- vs non-minimum-phase combs at matched magnitude — where any depth-only rule is at chance — it scores 90/90 at the operating point;
• on the boost history of one car (ours), it reproduces 7/7 real outcomes;
• its thresholds are provisional and fitted on that one car, which is exactly the kind of empirical constant you refuse to accumulate — and you are right to.

What would settle it is not more analysis on our data. It is one measurement session in a cabin the algorithm has never seen, with a ground truth that does not depend on anyone's tuning: sweep each driver solo at the listening position P0, then again with the microphone 20–30 cm to the side (P1). A dip that stays at the same frequency in both is a mode — EQ-able. A dip that moves is positional interference — not. That gives every dip a label a human can check, independent of the algorithm, and then the algorithm's single-point verdict either matches or does not. Two mic positions are the whole trick; a printed head with mics in both ears gives them for free. We have the protocol written up (~40–60 min a car, plain REW, no special gear) and can send it as a document.

Your archive almost certainly cannot answer this — I assume a single mic position per set — which is precisely why I am describing the protocol rather than asking for files.

2. What the body does — and this is the part I care about most

The question underneath the one above is bigger, and nobody in this hobby has data on it: how much of a car's response is the installation, and how much is the body? Concretely, the things that would change how people build systems:

• the ratio of the first arrival to the reflection cluster behind it (your TransferIrDiagnostics.MeasureCompactness and EstimateIrStart already measure exactly this) as a function of where a driver sits and where it aims — doors versus kick panels versus a dash firing at the windscreen;
• what fraction of a driver's band is non-minimum-phase at the listening seat, i.e. the part of the response no EQ, no crossover and no all-pass will ever fix. That single number would let two installs be compared honestly, before a single filter is dialled in;
• which cabin modes are the body's own (they should reappear in every car of the same shape, whatever is installed) versus the install's (they move when the driver moves).

That third one is why one cabin can never answer it, and two barely can.

The strongest evidence would be an A/B of placement in one cabin: the same driver measured solo at the same seat before and after it is moved or re-aimed — a door woofer versus the same driver in a kick panel, a mid on the A-pillar versus on the dash. Everything else stays; the difference is the placement. If you have ever moved a driver and re-measured, those two sets are worth more than a hundred single-position captures.

3. What we bring

• A cabin fully published under CC BY 4.0 — autosound-measurements: a VW Passat B8 saloon, eight channels of loopback-referenced IRs in your v7 format (raw and protective-HPF-compensated with your own code), the moving-mic RTA next to them, two DSP states the ear accepted, and the install and rig described in CAR.md.
• Tooling in Python that reads both worlds (REW → v7, and your files back), plus the excess-phase gate, a joint-phase summation metric and a jitter-robustness score for near-tie alignment candidates.
• Whatever we learn from your cabins, back to you, with the working shown — and gladly as a PR to Resonalyze if a result turns out to be worth encoding (your call whether it belongs, as with the boostability criterion).

4. What I am asking

1. Is the question interesting to you? If yes, the simplest form of joining forces is a shared, licensed pool of cabins with real metadata, which we both already half-have.
2. Do you have data that fits? The three cabins behind your heuristics, the newer sets you have not published, and above all any before/after of a driver being moved. Even single-position sets with the car and the placement written down move the body-versus-install question forward.
3. If you ever measure a new car, adding the second mic position (P1, 20–30 cm) costs one extra sweep per driver and turns the whole session into ground truth.

No obligation, and nothing here is urgent. But your TODO.md says several thresholds are anchored on your own cabins, and ours says the same about ours — that is the same wall from two sides, and it is a solvable one.

[DIMOSUS]

I'm getting a bit of information overload here :)

I can easily run this experiment in my own car and share the data with you. I can fix the microphone to the headrest and vary its position by changing the seat-back angle or simply sliding the seat forward and backward. I can also complement those measurements with near-field measurements of each driver.

One thing I'm still wondering about, though: how reasonable is it to rely on EQ boost in practice? My understanding has always been that boosting should be used rather cautiously, especially for dips, because you can quickly burn through headroom and driver excursion while getting very little actual improvement.

I understand that this is essentially what you're trying to distinguish here — dips that are actually boostable from those caused by cancellation. So I'm definitely interested in seeing whether the experiment can identify that reliably

[ayukhno]

Fair — that was a lot at once. Let me answer the one question you asked, because it is the important one, and leave the rest until there is data.

"How reasonable is it to rely on EQ boost?"

Mostly it is not, and I am not arguing for more of it. The gate's job is the opposite one: it is a reason to withhold boost, not a licence to apply it.

Our own practice is cut-first — gain balance is cut-only, and the mid in this car has ±1.1 mm of Xmax, so a boost anywhere near its bottom end is off the table regardless of what any criterion says. What the gate changes is the small set of cases where a tuner (or an auto-tuner) is about to boost anyway, because the dip looks like a resonance and the target curve asks for it. Two things can be under that dip:

• a resonance or absorption — minimum-phase, the inverse filter genuinely refills it, and the cost is only the energy you put back;
• a cancellation — two arrivals subtracting at the microphone. The filter refills nothing. You spend excursion and headroom making two arrivals cancel a little louder, and the null does not move.

The second case is where the headroom actually goes, and it is invisible on a magnitude plot: the two look alike. So the practical value of the criterion is that it says don't — and an auto-tuner that suppresses boost where the dip is non-minimum-phase is strictly safer than one that decides from magnitude alone, whatever the thresholds turn out to be. That is also why I proposed it as an optional criterion rather than something that unlocks new behaviour: it can only take boosts away.

I owe you a comparison on our own cabin between what your boost mask allows and what the gate allows, channel by channel. I have a first pass, but it was computed with a level bug in my harness (I had pulled REW's impulse responses peak-normalised, so every relative level in that run was wrong) — I found it afterwards, the published v7 files are unaffected, and I am re-running the comparison before quoting any numbers at you.

Your car

Yes, please — and what you describe is exactly the capture, not an approximation of it. Headrest-mounted mic, position varied by seat-back angle or by sliding the seat, is the two-point test; near-field per driver is the part I would have asked for next, since it separates what the driver does from what the cabin does to it.

The protocol is written up here, with the reasoning behind each constraint rather than just a list:

https://github.com/ayukhno/autosound-measurements/blob/main/protocols/mic-shift-out-of-sample.md

The short version: each driver solo, EQ and all-pass off (an all-pass is added excess phase by construction — it corrupts the one quantity being measured), protective high-passes fine and declared, sweeps on the timing reference, two mic positions 20–30 cm apart, one MMM, and one repeat of the first position untouched so the drift floor is known. Note how far the mic moved and in which direction. Resonalyze's own files are the ideal format — you would be measuring in your own app, at 4 repetitions, with coherence.

Two positions on your cabin is what turns this from "a criterion fitted to one car" into something with an out-of-sample answer — including the answer "no, it does not generalise", which is a perfectly good outcome and cheaper to find out this way than in someone's tune.

[DIMOSUS]

Before I spend time repeating this experimentally in my own car, I remembered that closely related work on spatial sampling in car cabins already exists and is probably worth mining first:

https://www.minidsp.com/applications/car-audio/1-microphone-techniques-car-cabins

The miniDSP article compares fixed-point measurements, moving-mic measurements, arrays and binaural approaches in car cabins. More importantly for this discussion, it points back to Earl Geddes' "localized sound power" work: six spatially distributed microphone positions in the listener ear-ellipsoid, averaged to improve resolution/reproducibility and correspondence with perception. It also links later work on higher-order automotive arrays.

This is not necessarily a direct answer to the exact excess-phase/boostability criterion we are discussing, but the underlying question — how strongly a local feature depends on microphone position, and how much spatial sampling is needed before treating it as a property of the system rather than a point cancellation — has clearly been studied before.

Could you please follow the references from that article (especially the Geddes/AES work and the later automotive-array paper) and check whether they already provide useful experimental results or methodology for our question? I would rather reuse existing evidence and only run a new car experiment for whatever remains genuinely unanswered.

[ayukhno]

Followed both trails. You were right that this is studied — and the existing work answers half our question quantitatively, while leaving the other half untouched in a way that I think sharpens what an experiment would have to do.

What the literature settles

Geddes & Blind, "The Localized Sound Power Method", AES Convention 76, paper 2127, 1984 (Ford Motor Company). The abstract states the finding in one sentence:

Single point microphone measurements in enclosed spaces lack sufficient stability to be useful for sound system equalization (above the Schroeder frequency).

Note the bound — above the Schroeder frequency.

Attribution, so nobody builds on sand: that quotation is the abstract, which is public on the AES page; the paper itself is behind $33 and I have not read it. Everything else usually credited to it — the six microphones, the ear ellipsoid, the ~1/3-octave reliable resolution — I have only from miniDSP's summary, not from the source. The abstract states the problem and that a technique was developed; it states no numbers. If you have E-Library access, the six-point geometry and what it actually buys are worth reading first-hand before either of us designs around them.

Strauß & Treichel (Fraunhofer IDMT) with Kessler (BMW), DAGA 2010 — the higher-order array paper — supplies the number and the mechanism:

In a small enclosed space, e.g. inside a car, the Schroeder frequency is about 150–200 Hz […] For pure tones above the Schroeder frequency more and more room modes are excited. The result of many virtually random modes being added will be a Gaussian distribution in both the real and the imaginary parts […] resulting in a Rayleigh distribution of sound pressure levels. This accounts for the much higher likelihood of measuring a sound pressure value that is too low than too high.

That last clause is the part I did not expect and think matters most to us. Above ~200 Hz in a car, a deep null at a single point is the statistically expected outcome of a random modal sum, not evidence about the system. Peaks are rarer than dips by construction. So "how much spatial sampling before a local feature counts as a property of the system" has an answer with a frequency attached: below ~150–200 Hz one point is defensible; above it, one point is a draw from a Rayleigh distribution and the field standard is the ellipsoid average.

The miniDSP note you linked lines up with that from the practitioner side: single fixed point, reproducibility 2/5 and errors up to ~20 dBr concentrated in the mid and high frequencies; moving mic 4/5; array 5/5.

Our own data measures that boundary, by accident

We re-measured the whole front yesterday on a tripod, and the repeat captures put numbers on exactly this transition — same channel, same position, minutes apart:

band	repeat-to-repeat difference
25–150 Hz	0.19 dB RMS
150–500 Hz	0.46 dB RMS
500 Hz – 1 kHz	3.01 dB RMS
1–2 kHz	1.28 dB RMS
2–5 kHz	2.87 dB RMS

That is the Schroeder boundary showing up as reproducibility, in a real cabin, without anyone looking for it. And the two features I described in #88 sort themselves by it: the left woofer's +10 dB at 100–125 Hz and −9 dB at 160 Hz sit below the boundary — discrete, sparse, position-stable, and a legitimate target for a filter. The mid's 736 Hz dip and the right woofer's 2.2 kHz instability sit above it, in the Rayleigh regime.

What the literature does not answer

All of that work treats position dependence as noise to be averaged away. It answers what should I equalize to — the ellipsoid average. It does not answer which features of a measurement are correctable at all, because averaging deliberately destroys the very thing that distinguishes them: whether a dip moves when the microphone moves.

That distinction is the whole subject here. A mode and a cancellation both look like a dip in a six-point average. They differ in that one stays put and the other does not, and no amount of spatial averaging recovers that — averaging is the operation that throws it away.

So I would say: the sampling question is answered, the classification question is not. And the two are not in conflict; the first tells us the conditions under which the second is even meaningful.

Three things I would change on the strength of this, before anyone measures anything

1. Frequency-scope the criterion. Our excess-phase thresholds were fitted without this distinction. A single-point deep null above ~200 Hz should never be certified as a system property, whatever the criterion says about it — that is now a statement with a citation behind it rather than a hunch.
2. The Rayleigh asymmetry is an independent argument for cut-first tuning, and a much better one than "boosting burns headroom": above the Schroeder frequency a measured dip is more likely to be a sampling artefact than a measured peak is. That is a physical reason to distrust dips specifically, which is exactly what your question in this thread was about.
3. The capture protocol should become the ellipsoid, not two points. Our mic-shift-out-of-sample.md asks for P0 and P1 20–30 cm apart. Geddes' six points in the ear ellipsoid is both the field standard and a superset of what the stay-vs-move test needs: six positions give the average the literature trusts and the per-position spread that labels each dip. I will rewrite the protocol that way before asking anyone to run it — including before running it in our own car.

So: please do not spend a session in your car on the old two-point version. If you still want to run it after that rewrite, it will be a better experiment; and if the ellipsoid spread turns out to answer the classification question on its own, it may not need a dedicated experiment at all.

Thank you for the redirect — this is the second time this week that checking before measuring has changed what the measurement should be.

[DIMOSUS]

Direct link to paper https://pub.dega-akustik.de/DAGA_2010/data/articles/000235.pdf

[DIMOSUS]

And one more research on this topic https://www.linkedin.com/pulse/determining-repeatability-automotive-audio-testing-pt-soendergaard-mfzwe/

[ayukhno]

Ran it in our own car before asking you to run it in yours — which is what I should have done before proposing it at all.

The pilot

One channel (left A-pillar midrange), nine sweeps, 97 seconds, microphone hand-held — because a tripod cannot be repositioned six times inside a cabin, and pretending otherwise is how a protocol gets published and never run. Order: centre, three left positions, centre, three right positions, centre. The three centres are the noise floor; the six are the signal.

Single sweeps, not averaged: four sweeps over 23 seconds average four slightly different positions when the microphone is in a hand, which smears exactly what the test reads.

The dip near 800 Hz:

	where the dip sits
centre ×3	800 · 818 · 800 Hz — scatter 1.0 %
six positions	688 · 760 · 772 · 789 · 806 · 925 Hz — scatter 9–14 %

The same hand, returning to the centre, puts the dip back within one percent. Moving it 18 cm moves the dip by up to 390 Hz. Signal to noise, an order of magnitude.

Level spread told the same story more weakly — six positions scatter about twice as much as three centres in every band — but the frequency of the dip is the discriminating measurement, and it is not marginal.

What that settles

The stay-versus-move test survives a hand, provided the capture includes repeats at one position to establish the floor. Without those anchors the six numbers mean nothing; with them, they mean a great deal. That is now a requirement in the protocol rather than a suggestion.

And it answers your question in this thread with our own data. That dip is a cancellation: it moves. Boosting it would spend excursion to make two arrivals cancel a little louder. It is also, independently, above the Schroeder frequency — and the June session had already hinted at it from the other side: opening the car's windows halved the 180 Hz hump and left the 736 Hz dip unmoved, which is a mode and an interference behaving exactly as they should.

So on this cabin, three independent lines agree, and the practical rule falls out of them rather than out of a criterion: below ~150–200 Hz a single point is worth trusting and a dip may be worth a filter; above it, a single-point dip is not evidence about the car.

What this means for your session

Worth running, and the design has changed since I first described it:

https://github.com/ayukhno/autosound-measurements/blob/main/protocols/mic-shift-out-of-sample.md

Six positions in the ear ellipsoid rather than two, kept separately rather than averaged — the average is what the field equalises to, but it is also precisely the operation that destroys the stay-versus-move signal. Anchor repeats at one position, mandatory. Hand-held is acceptable and its cost is stated (1–2 dB RMS per position); a bar with six marked holes removes that cost for the price of an afternoon and makes the positions reproducible between cars, which hand-holding never will.

If it must be hand-held, take more positions rather than fewer: precision per sample is what a hand loses, and count is what it can cheaply buy back.

What remains genuinely unanswered

The sampling question is settled by the literature you pointed at. The classification question is not, and cannot be by averaging: a mode and a cancellation are identical in a six-point average and differ only in the behaviour that averaging discards. Nothing in Geddes or the automotive-array work asks which features of a measurement are correctable at all.

And our own cabin cannot settle it either — the criterion's thresholds were fitted on this car, so ground truth from it is in-sample and carries almost no weight. That is the one thing only a second cabin provides, and it is why your car is still worth an hour whenever you feel like spending one.

Leaving this open rather than closing it, since that hour is still on the table.

[DIMOSUS]

This topic has turned out to be much more researched than we expected:
https://testgear.audiofrog.com/wp-content/uploads/2023/06/A-Straightforward-Stereo-Tuning-Process-and-Some-Notes-About-Why-it-Works_V2.pdf
https://www.audiosciencereview.com/forum/index.php?threads/moving-mic-vs-averaged-vs-single-point-measurements-which-is-better.43880/
https://www.audiofrog.com/measuring-frequency-response-single-mic-or-multi-mic-average/

[ayukhno]

Before the substance, so you read it in the right light. What follows is a literature review done with
AI, not my own analysis. It is past my depth — I cannot defend these workings line by line, and I would rather
say so than let them stand as my judgement. I am passing them on because the sources you gave deserved to be
read properly and someone should set down what is in them.

And to be straight with you: I am stepping back from this line. Take what is useful, ignore the rest, and
treat none of it as a claim I am in a position to argue for.

---

Read all three, and the paper the ASR thread points to. They change more than I expected — including one
correction to what I wrote above.

The correction first

I cited Geddes & Blind, AES Convention 76, paper 2127, 1984 and said honestly that I had only the
abstract. The ASR thread names the right source: there is a peer-reviewed Geddes & Blind, 1986, and the
1984 item is the convention paper. Anyone building on this should go to the 1986 one; my citation was the
weaker of the two.

What Wehmeyer's article adds: the same boundary, measured in a car

His single-vs-multi-mic article is the closest thing to a direct answer either of us has found, because it is
in a car and about a car. Five positions in a 7-inch circle around the driver's head, one side's speakers:

band	what he sees across the five positions
below 200 Hz	"the five measurements are very similar"
200 Hz – 1 kHz	transition, "measurements begin to diverge"
above 1 kHz	"moving the mic changes the frequency response by 25 dB or more"

That is the Schroeder boundary again, from spatial spread rather than from theory, in the right kind of
enclosure. It brackets what our tripod repeats showed as reproducibility (0.19 dB RMS at 25–150 Hz, 3.01 dB
RMS at 500 Hz–1 kHz) — his is position-to-position, ours was repeat-to-repeat at one position, and both put
the transition in the same place.

And he converts it into something more useful than a frequency limit. A single-mic measurement smoothed
to 1/6 octave tracks the five-position average closely; where it still disagrees, the disagreements are
narrow. So his rule is not "measure more", it is:

we shouldn't use a filter with a Q value higher than 6

That is a bandwidth ceiling rather than a frequency ceiling, and it is better than what I proposed. My point
1 above was "frequency-scope the criterion". His rule says the real predictor of "this feature is not a
property of the system" is how narrow it is, at any frequency — which is exactly what makes a feature
position-dependent in the first place. I would rather adopt that than my own version.

MMM is out, on the MMM literature's own terms

The ASR thread's best reference is Jean-Luc Ohl's MMM paper. It is an advocacy paper for MMM, and it excludes
our use case in its own words:

MMM is quick and efficient for equalization of a loudspeaker but because it is missing the time and phase
information, it is not a tool to set up crossovers, time align speakers, ... (for this, you need first an
MLS or sinesweep measurement at one point).

Its validity conditions rule out a car twice over: a 5 m scanning path, a scanned volume about 1/5 of the room
dimension, the microphone no closer than 1 m to any boundary, and the round-robin result it rests on
(Critchley & Dunbavin) is for rooms of 30 m³ or more. A cabin is roughly 3 m³ and has no point 1 m from a
boundary.

A poster in the thread states the same split more directly, and it is the cleanest statement of it I have seen:

If you measure all channels together and apply the same EQ to all channels: use MMM because you don't need
any phase info. If you measure and EQ each channel individually, use single point. The phase info is
necessary to simulate and correct L/R/Sub summation.

That is Resonalyze's case exactly. So the tension I thought existed between the Geddes line and what we both
do is not a tension: the two literatures optimise different objectives. Spatial averaging answers what should
I equalise to. Per-driver single-point answers what are these drivers doing and how will they sum — and it
is the only one of the two that can, because averaging discards the phase.

The thing I did not have: a second classification axis, for free

Wehmeyer's tuning document has a test I had not considered, and it costs nothing:

look for big dips in the response that don't occur when only one channel is playing. These dips, if there
are any, are phase problems and unless they are caused by an error in your delay settings, you can't fix
them. If there's a big peak (more than 6 dB), you can remove it by equalizing both channels equally.

Mute a source instead of moving the microphone. That separates inter-channel cancellation with certainty
and no mic movement at all — it is a different axis from the stay-versus-move test, not a substitute. Ours
asks whether a feature belongs to the room or to the geometry between the driver and this point; his asks
whether it exists only in the sum. Both are needed, and his is free for anyone measuring per-driver, which
Resonalyze already does.

His practical corollary is worth quoting too, because it is about the image rather than the curve: if a dip on
one side cannot be equalised, put a matching dip in the other side — degrade the response to preserve the
centre image.

A third class of dip that neither of our criteria has

Pages 28–30 of his tuning document describe a case that is neither a mode nor an interference null between
channels: a crossover placed where the woofer has already gone directional. The direct sound through the
crossover is fine; the reflected sound has a hole there, because the driver is no longer radiating into
the room at those angles. The microphone sums both and shows a dip. Equalise that dip flat and the direct
sound — the part that carries the image — becomes peaky and bright.

That matters to us because our gate would likely refuse to boost it, but for the wrong reason, and would say
nothing about the actual fix, which is not a filter at all: choose the crossover from the driver's
directivity. He gives a diameter-versus-frequency chart for it.

So the taxonomy is at least: modal (position-stable, EQ-able below the transition), positional
interference (moves with the microphone), inter-channel cancellation (present only in the sum), and
power-response deficit (direct fine, reflected holed, fix the crossover). Our excess-phase gate was built
to split the first two. It does not know about the last two, and two of the four have tests that are cheaper
than anything we have been designing.

What this leaves

Less than I thought, and that is the useful outcome.

For a practitioner, the combination of the 200 Hz / 1 kHz boundary, the Q ≤ 6 rule, the mute-one-channel test
and the directivity chart already decides most real cases without any criterion of ours. I would rather say
that plainly than defend the gate's scope: its remaining value is in the transition region and on broad
features — where the simple rules are silent and where a wide, deep, position-stable dip might genuinely be
worth a filter or might be an excess-phase artefact.

That is a much smaller claim than the one I opened this thread with, and it is what a second cabin would
actually have to test. The protocol would want rewriting against the Q ceiling rather than only against a
frequency scope — but per the note at the top, I am not the one who will be doing that, so please do not hold
a session in your car waiting on it.

Thank you for the three links. Three times in this thread now, reading first has changed what the measurement
should be, and twice it has made the measurement smaller. That seems the most useful thing to leave here.

[DIMOSUS]

I think we are mostly converging, but I would draw the boundary a little differently.

For crossover integration, delay, polarity, and phase alignment, I agree that the measurement must remain a single fixed reference point. That is the dataset Resonalyze uses to preserve the relative timing and phase between drivers, and spatial averaging would destroy exactly the information needed for complex summation and alignment.

For magnitude EQ, though, the situation is different. Time reference and inter-position phase are not important there. What matters is obtaining a representative frequency response over the listener's head area. For that purpose, both MMM and a multi-microphone array are entirely valid tools. I therefore would not say MMM is "out" for the Resonalyze workflow — it is simply unsuitable for the timing/phase part of it.

At the same time, Wehmeyer's results suggest that the existing single-point approach may already be considerably better than raw single-point measurements would imply. His measurements show very large point-to-point differences at high frequencies, but after roughly 1/6-octave smoothing, a single position follows the five-position average quite closely, with the remaining differences being mostly narrow features. His practical Q < 6 recommendation is important here.

That maps surprisingly well onto what Resonalyze already does:

single fixed point → psychoacoustic smoothing → avoid boosting dips → conservative EQ bandwidth

If I add a Q ≤ 6 limit for automatic car EQ, then most of the highly position-dependent narrow structure is intentionally excluded from the optimization.

My current engineering estimate — and I want to stress that these are my estimates, not numbers claimed by the cited papers — would be roughly:

• conservative single-point EQ with psychoacoustic smoothing, no dip boosting, and Q ≤ 6: around ±2–3 dB spatial uncertainty;
• a well-spaced 5–6 microphone average: around ±1–2 dB;
• larger arrays or a good MMM measurement may approach roughly ±1 dB for the broad tonal response.

The exact values will obviously depend on frequency, microphone spacing, the vehicle, and the installation. I would not expect a 6-mic array to provide a √6 improvement at low frequencies because the measurement positions there are strongly correlated.

So the practical difference may be smaller than it first appears. A raw single-point response can be wildly misleading, but a deliberately bandwidth-limited single-point EQ is a very different thing. The array then improves an already reasonably robust result rather than rescuing an unusable one.

Where I think a multi-microphone array becomes especially interesting is that it can preserve more information than MMM. Instead of keeping only the spatial mean, Resonalyze could retain both:

• the spatial average μ(f) for EQ;
• the position-to-position spread σ(f) as an indication of how spatially stable each feature is.

A +4 dB feature with σ = 0.5 dB is a very different EQ candidate from a +4 dB feature with σ = 4 dB, even though their spatial averages may look identical.

So at this point I would separate the problem into two datasets rather than choose one measurement technique for everything:

Reference IR: one fixed microphone position for crossover, delay, polarity, and phase.

Spatial FR: MMM or multiple microphone positions for magnitude EQ and spatial-stability information.

That seems to preserve the strongest part of both approaches.

I plan to take measurements in my car at 6 points for each of the speakers individually to confirm my assessment.

[DIMOSUS]

New data is up in Resonalyze-test-data (CC BY 4.0), both sets described in the repository README.

v8 — seven microphone positions in one cabin. This is close to the ground truth you describe in §1, and it went past two positions. In the BMW F30 already published as v5, the left midrange and the left tweeter were each swept solo eight times while a single microphone was physically carried through seven positions around the listening seat — centre, then left / left forward / left further forward and right / right forward / right further forward — and then returned to the centre for a closing repeat. 96 kHz / 24 bit, loopback-referenced, two averaged runs per position. The closing centre repeat is deliberate: it is the same position as the opening one, so it shows the run's own repeatability before you read anything into a difference between positions.

A dip that holds its frequency across those seven and a dip that walks are separable there without reference to anyone's tuning, which is what you were after. Two caveats. It is two drivers, not the whole system. And the geometry is not in the files — the positions carry direction labels only, no distances; ask if you need the spacing. The tweeter runs sit behind a protective 48 dB/octave Butterworth at 1 kHz and carry the compensation term, the midrange runs have none.

mid MMM.json and tw MMM.json sit alongside them: moving-microphone captures of the same two drivers, so the seven-position average has something to be judged against.

v6 — moving-microphone averages of the same car. One MMM curve per channel for all seven channels of the v5 system, captured with the DSP bypassed, plus the tuned Virtual DSP project that pairs each swept measurement with its MMM curve for the hybrid view.

Nothing here touches §2, to be clear: the F30's drivers sit in their factory locations, and none of them was moved and re-measured.

[DIMOSUS]

Three additions since the last comment, all pushed to Resonalyze-test-data.

The v8 spacing, since I said to ask. The side positions sit 10 cm to the left and to the right of the centre; the forward pair adds 10 cm forward to that, the further-forward pair 20 cm. The seven span 20 cm across and 20 cm deep — as wide as the tripod would reach. The capsule pointed at the ceiling at every one of them, so the drivers are at grazing incidence throughout.

That bears on your §1 protocol directly: you wanted 20–30 cm of lateral separation, and left-to-right here is 20 cm, at the bottom of that range. Centre against a side position is only 10 cm.

One trap before you compare the seven sweeps to the MMM. An impulse response in this format is always stored raw — no calibration is ever baked into one — while a live capture bakes its correction into the drawn curve and records it in calibrationCorrectionDb. The curve in force through that session was the 0-degree on-axis one. Below 4 kHz it makes no difference; by 10 kHz it is worth 1.3 dB and by 20 kHz 9.7 dB. So read the sweeps with that curve selected, or take the correction back out of the MMM curve, before putting one against the other. Read that way, the average of the seven positions and the MMM agree in tilt across the band. It is all in the README now.

And a nearfield/ set, which is not what you asked for but may be closer to §2 than anything else I have. The seven drivers of the root dataset, measured a second time with the microphone right at each cone, on axis, ten minutes after the listening-seat pass and on the same rig — 44.1 kHz, four averaged runs, the same sweep. Six files carry the root's names; woofer.json is the subwoofer.

It is not the placement A/B you called the strongest evidence: the microphone moved, not the driver. But both of your §2 metrics need a baseline no seat measurement can supply. To say what fraction of a driver's band is non-minimum-phase at the seat, you need the driver's own excess phase to hold it against; to read compactness as a function of placement, you need the reflection-free end of the same scale for the same driver in the same install. This gives both, for every driver of one car whose seat measurements you already have.

Its limits, so it is not over-read: a microphone at the cone inside a car is not anechoic — the door cavity, the mounting and the boundaries are all still in it. It is the driver as installed with the reverberant field suppressed, not a free-field response, and it is a true nearfield only while capsule-to-cone stays small against the wavelength and against the radiator: the low end for the woofers, close-mic on axis rather than a gated curve for the tweeters.

One thing fell out of assembling it that you can use as it stands. Comparing the two passes' impulse-response peaks: for the mids and tweeters the peak moves by the equivalent of 0.77 to 1.31 m, plausible driver-to-head paths. For the right under-seat woofer and the subwoofer it moves by 4.36 and 4.08 m — no direct path inside a car is that long. At the seat those two peaks sit on the reflection cluster rather than on the direct arrival. A labelled pair of examples of exactly the confusion your gate exists to avoid, with the same driver's clean arrival sitting next to it.

[ayukhno]

A follow-up from this thread has its own issue, because it outgrew a comment: #178.

Your redirect to Geddes & Blind changed how we capture, and your v8 set turned out to answer the "one car" objection to the result — one microphone through seven positions makes three pairs 20 cm apart, which is the same shape as the three ear heights we measured. The question at the end of it is about whether we are reading your prior art correctly or over-reading it.

Nothing else here changes; this thread's own business is untouched.