Before the technical part, so it frames the rest.
The idea is not mine. I saw it at flatlinertuning.com/stereo-image.php — a free browser tool that loads a track and shows where the mix places each frequency between left and right: every row a frequency, left-to-right the stereo position, brightness the level, the width of the glow how wide that frequency is spread. That page is careful about its own scope: it says outright that it shows you the recording, not your car, and is therefore the reference you compare a tune against.
What I wanted was the missing half — read the same picture from microphone captures in the car, so a tuner could put the two side by side: where the mix puts a sound, and where the listener actually hears it.
I could not get that to a state that is useful for tuning. This is how far it got, which is why it arrives as a finding and not as a method.
My hypothesis in sending it is simply that you might find it interesting, or that you — knowing the mathematics of sound far better than I do — may see a way to make it useful that I could not see. That is a guess on my part, nothing more.
If this pulls you away from your own work, or is redundant, or is just not interesting to you: please ignore it, or leave it for whenever, or for never. I have no expectation of any action from your side. If it turns out to help, I will be glad, and that is the whole of it.
And if any of it needs testing: I am happy to run experiments in the car. Ask for whatever measurement would settle a question, and I will make it.
Thank you.
Below is the work itself. A negative result, with the prediction written down before the listening and all of it published. One section is about your data rather than mine, and there is exactly one question at the end.
What was asked
Can a per-channel DSP steer the interaural cues a listener uses to place a stereo image? We measured a cabin at two microphone positions at ear locations (±9 cm), repeated at three heights (±7 cm), on a tripod that never moved.
The model is not the problem. Predicted against measured: 0.39 dB and 9 µs, and impulse peaks reconstructed from the prediction land within 10 µs of the measured ones with no fitting. From one raw two-ear session you can compute at the desk what a DSP change will do to interaural cues, and it comes true.
What fails is what the numbers mean.
The scissors
What survives the ear moving is not steerable. For one driver the interaural ratio is H_d(right) / H_d(left) — the processing chain and the channel delay cancel exactly. A per-channel edit cannot move its own driver's interaural cue; it only changes which driver dominates in which band of the sum.
What is steerable does not survive the ear moving. σ of the summed cues across the three ear pairs:
| band |
σ ITD |
σ ILD |
sign |
| 500 Hz |
60 µs |
0.64 dB |
holds in both |
| 1 kHz |
235 µs |
4.03 dB |
ILD flips |
| 2 kHz |
383 µs |
2.56 dB |
both flip |
| 3.15 kHz |
555 µs |
2.36 dB |
ITD flips |
| 5 kHz |
420 µs |
1.25 dB |
ITD flips |
Above ~500 Hz the intersection is empty.
The prediction, and the ear
We wrote down beforehand what one specific move — a −3 dB peaking cut at 420 Hz on the virtual front-right channel — should do, with the refutation criteria for each of three predictions. Then the car's owner listened, on a competition evaluation disc, with the hardware unchanged.
| prediction |
ear |
| the centre in 300–600 Hz moves left |
no — the images stretched apart, the centre did not travel |
| the top does not move |
yes, confirmed |
| focus improves, ρ 0.78 → 0.83 |
no, the opposite |
The survivor is the interesting one: the model's band selectivity was confirmed by ear — a cut at 420 Hz stayed in its band and did not drag the top with it. What failed is the next link, that a change in band level moves the position of an image.
The part that concerns you directly
The obvious objection is "one car". While writing this up we realised you had already published the data that tests it, so we ran our analysis on your set instead of asking you to run ours.
Resonalyze-test-data v8 carried one microphone through seven positions in the BMW F30 with the left midrange swept solo. That makes (L,R), (LF,RF), (LFF,RFF) three pairs 20 cm apart with the pair translated forward — the same shape as our three ear heights, in a different cabin, on a different rig, taken by a different program.
The shape reproduces.
| band |
ITD per pair, 0 / +10 / +20 cm |
σ |
ILD per pair |
σ |
| 500 Hz |
−302 / −417 / −240 µs |
90 µs |
+0.29 / +2.15 / +3.24 dB |
1.49 dB |
| 1 kHz |
−260 / railed / −198 µs |
44 µs |
−2.14 / +3.64 / +1.87 dB |
2.96 dB |
| 2 kHz |
−146 / +135 / −271 µs |
208 µs |
+3.70 / +1.30 / −2.36 dB |
3.05 dB |
| 3.15 kHz |
railed / −188 / −323 µs |
96 µs |
+1.61 / +0.54 / −2.17 dB |
1.95 dB |
| 5 kHz |
railed / +292 / −83 µs |
265 µs |
−1.05 / +0.27 / +1.86 dB |
1.46 dB |
500 Hz is the only band where neither cue changes sign across the three pairs. Above it the level cue flips in four of five bands and the delay in three, and the top two rail out of the ±583 µs geometry for a 20 cm spacing — interference between the two points, not an arrival difference, so they are excluded from σ rather than averaged in.
Your own closing repeat sets the noise floor, which is why the spread can be read as real: the opening and closing centre sweeps of the same position differ by 10–21 µs and 0.12–0.41 dB across all five bands.
Two cabins is still two. But it is your cabin, and the script is published, so you can check whether we read your set correctly. The tw half of v8 would extend this above 5 kHz and we did not download it — that one is yours if you want it.
The one question
In #91 you sent us to Geddes & Blind, and through them to the Schroeder-frequency argument: above it a single measured point in an enclosed space is unreliable, and in a cabin that boundary is 150–200 Hz.
Our result looks to us like the same boundary seen from the other side — not "one point is unreliable" but "the difference between two points is unreliable in the same regime, and that difference is the whole stereo cue". Is that a fair reading of the prior art you pointed at, or an over-reading of it?
That is the only thing this issue asks of you.
If it is a reading instrument rather than a control surface
That is where we ended up, and it reads as a loss but is really a specification. If a tool wants to describe a stage from a two-ear capture rather than steer one, our σ table says how much it may claim: a position below 500 Hz, a direction and not an amount at 1 kHz, and nothing stable above that. Two things follow that are cheap to build and hard to guess — show a zone above 500 Hz rather than an arrow, and weight a per-band diagnosis by the stability of that band's cue. On our tuned preset 32 of 34 bands cross the starting thresholds: a diagnostic that flags almost everything says nothing.
Neither of those needs our conclusion to be right. They only need the σ table, and the σ table is measurement.
This is also the honest answer to where the idea came from. The tool that prompted it draws its line at the recording and says so; the σ table is how far the line can be moved toward the car, and no further.
Everything is published
autosound-measurements → cars/vw-passat-b8-sedan-lhd/2026-09-03_two-ear-stage/ — CC BY 4.0.
The full note, the pre-registration exactly as written before the listening (with the Ukrainian original verbatim so the translation can be checked), the acceptance criteria as JSON, the band-by-band stage picture, the raw 24.6 MB REW session, the σ chart, and v8-crosscheck.py, which fetches your data itself.
Two things about the note are deliberate. Our own reading is one paragraph, at §6, kept short and marked as ours — sections 1–5 are inputs, and we would rather have your reading before ours colours it. And §8 argues against us: there is no manikin on this rig, so an "interaural level difference" here is cabin interference between two bare points 18 cm apart rather than head shadow plus pinna filtering, and a manikin might well show a more stable level cue than we did. That is the strongest objection to our conclusion, and we would rather state it than have it found.
One car, one listener, one evening, plus three pairs in yours. Treat it as that.
Before the technical part, so it frames the rest.
The idea is not mine. I saw it at flatlinertuning.com/stereo-image.php — a free browser tool that loads a track and shows where the mix places each frequency between left and right: every row a frequency, left-to-right the stereo position, brightness the level, the width of the glow how wide that frequency is spread. That page is careful about its own scope: it says outright that it shows you the recording, not your car, and is therefore the reference you compare a tune against.
What I wanted was the missing half — read the same picture from microphone captures in the car, so a tuner could put the two side by side: where the mix puts a sound, and where the listener actually hears it.
I could not get that to a state that is useful for tuning. This is how far it got, which is why it arrives as a finding and not as a method.
My hypothesis in sending it is simply that you might find it interesting, or that you — knowing the mathematics of sound far better than I do — may see a way to make it useful that I could not see. That is a guess on my part, nothing more.
If this pulls you away from your own work, or is redundant, or is just not interesting to you: please ignore it, or leave it for whenever, or for never. I have no expectation of any action from your side. If it turns out to help, I will be glad, and that is the whole of it.
And if any of it needs testing: I am happy to run experiments in the car. Ask for whatever measurement would settle a question, and I will make it.
Thank you.
Below is the work itself. A negative result, with the prediction written down before the listening and all of it published. One section is about your data rather than mine, and there is exactly one question at the end.
What was asked
Can a per-channel DSP steer the interaural cues a listener uses to place a stereo image? We measured a cabin at two microphone positions at ear locations (±9 cm), repeated at three heights (±7 cm), on a tripod that never moved.
The model is not the problem. Predicted against measured: 0.39 dB and 9 µs, and impulse peaks reconstructed from the prediction land within 10 µs of the measured ones with no fitting. From one raw two-ear session you can compute at the desk what a DSP change will do to interaural cues, and it comes true.
What fails is what the numbers mean.
The scissors
What survives the ear moving is not steerable. For one driver the interaural ratio is
H_d(right) / H_d(left)— the processing chain and the channel delay cancel exactly. A per-channel edit cannot move its own driver's interaural cue; it only changes which driver dominates in which band of the sum.What is steerable does not survive the ear moving. σ of the summed cues across the three ear pairs:
Above ~500 Hz the intersection is empty.
The prediction, and the ear
We wrote down beforehand what one specific move — a
−3 dBpeaking cut at 420 Hz on the virtual front-right channel — should do, with the refutation criteria for each of three predictions. Then the car's owner listened, on a competition evaluation disc, with the hardware unchanged.The survivor is the interesting one: the model's band selectivity was confirmed by ear — a cut at 420 Hz stayed in its band and did not drag the top with it. What failed is the next link, that a change in band level moves the position of an image.
The part that concerns you directly
The obvious objection is "one car". While writing this up we realised you had already published the data that tests it, so we ran our analysis on your set instead of asking you to run ours.
Resonalyze-test-datav8carried one microphone through seven positions in the BMW F30 with the left midrange swept solo. That makes(L,R),(LF,RF),(LFF,RFF)three pairs 20 cm apart with the pair translated forward — the same shape as our three ear heights, in a different cabin, on a different rig, taken by a different program.The shape reproduces.
500 Hz is the only band where neither cue changes sign across the three pairs. Above it the level cue flips in four of five bands and the delay in three, and the top two rail out of the ±583 µs geometry for a 20 cm spacing — interference between the two points, not an arrival difference, so they are excluded from σ rather than averaged in.
Your own closing repeat sets the noise floor, which is why the spread can be read as real: the opening and closing centre sweeps of the same position differ by 10–21 µs and 0.12–0.41 dB across all five bands.
Two cabins is still two. But it is your cabin, and the script is published, so you can check whether we read your set correctly. The
twhalf ofv8would extend this above 5 kHz and we did not download it — that one is yours if you want it.The one question
In #91 you sent us to Geddes & Blind, and through them to the Schroeder-frequency argument: above it a single measured point in an enclosed space is unreliable, and in a cabin that boundary is 150–200 Hz.
Our result looks to us like the same boundary seen from the other side — not "one point is unreliable" but "the difference between two points is unreliable in the same regime, and that difference is the whole stereo cue". Is that a fair reading of the prior art you pointed at, or an over-reading of it?
That is the only thing this issue asks of you.
If it is a reading instrument rather than a control surface
That is where we ended up, and it reads as a loss but is really a specification. If a tool wants to describe a stage from a two-ear capture rather than steer one, our σ table says how much it may claim: a position below 500 Hz, a direction and not an amount at 1 kHz, and nothing stable above that. Two things follow that are cheap to build and hard to guess — show a zone above 500 Hz rather than an arrow, and weight a per-band diagnosis by the stability of that band's cue. On our tuned preset 32 of 34 bands cross the starting thresholds: a diagnostic that flags almost everything says nothing.
Neither of those needs our conclusion to be right. They only need the σ table, and the σ table is measurement.
This is also the honest answer to where the idea came from. The tool that prompted it draws its line at the recording and says so; the σ table is how far the line can be moved toward the car, and no further.
Everything is published
autosound-measurements→cars/vw-passat-b8-sedan-lhd/2026-09-03_two-ear-stage/— CC BY 4.0.The full note, the pre-registration exactly as written before the listening (with the Ukrainian original verbatim so the translation can be checked), the acceptance criteria as JSON, the band-by-band stage picture, the raw 24.6 MB REW session, the σ chart, and
v8-crosscheck.py, which fetches your data itself.Two things about the note are deliberate. Our own reading is one paragraph, at §6, kept short and marked as ours — sections 1–5 are inputs, and we would rather have your reading before ours colours it. And §8 argues against us: there is no manikin on this rig, so an "interaural level difference" here is cabin interference between two bare points 18 cm apart rather than head shadow plus pinna filtering, and a manikin might well show a more stable level cue than we did. That is the strongest objection to our conclusion, and we would rather state it than have it found.
One car, one listener, one evening, plus three pairs in yours. Treat it as that.