Arguments on properly recording/sampling acoustic pianos (+more)

scherbakov.​al All three points are on target, but they're very different in terms of how "fixable" they are. Let me go through them one by one.

1. Differentiation — the most harmless one. Yes, the EMF is proportional to velocity, hence the +6 dB/oct slope. But this is a linear, stationary transformation: an integrator or a digital shelf filter can return the spectrum to normal without any loss of information. The cost is noise in the low register, and the fact that after correction you run into the pickup's own LC resonance (cable + preamp input), which no amount of EQ can remove from the signal, because it depends on cable length and load. So the problem isn't the derivative itself — it's that the timbre starts depending on the wiring.

2. Polarization — this, in my view, is the killer point, and it's more serious than "just quieter." A piano string has two planes of vibration with different decay times: the vertical one (perpendicular to the soundboard) is more heavily loaded by the bridge and dies out quickly — that's the loud attack; the horizontal one is weakly coupled to the bridge and lives on much longer — that's exactly the "aftersound" and the famous double decay. The hammer strikes vertically, and the horizontal component builds up afterward through coupling at the bridge and coupling between the strings of the unison.

A magnetic pickup positioned under the string reads mainly the vertical component. The horizontal one gives a weak signal, and at double the frequency at that, if the string passes symmetrically over the pole. Result: you record the loud onset and lose the long tail. Not "a slightly different timbre" — a fundamentally shortened, unnatural decay. And along with it, the unison chorus beating disappears too, because the shimmer of the unison lives mainly in that same long-lived plane.

3. Nonlinearity and inharmonicity. I agree on the mechanism, but I'd put the emphasis on intermodulation rather than harmonics. Harmonic distortion lands on exact integer multiples of each input partial, whereas real partials are stretched — and the higher the partial number, the greater the mismatch. Plus the derivative from point 1 boosts exactly the high end, where the stretching is greatest. And in a chord, sum-and-difference tones are added up from dozens of inharmonic components at once, and they don't land on any grid at all. That's what produces that "muddy and digital" quality. At soft dynamics the effect is small; at forte it isn't, because the nonlinearity grows with amplitude.

And I'd sign off on your conclusion as it stands: what you'd get is an interesting electroacoustic instrument with its own character — not a "cleaner" recording of a piano. And funnily enough, all three mechanisms listed produce exactly the metallic overtone that the whole scheme was supposed to eliminate. Except now it's born in the pickup instead of the string, and this time you can't fix it by voicing.

Reply generated by Claude AI

dikrek So by all means build it. It could be an interesting electro-acoustic instrument. But calling it a more accurate or “proper” recording of an acoustic piano is physically indefensible.

It is physically more accurate in comparison to mic recordings. It records as directly as possible the actual source of the note.
The overall construction, especially the nasty frame and undampened strings\aliquots resonance is what is eliminated here, for good. Mics capture all that nonsense. so those have to be properly treated acoustically for accurate sound representation.

Of course it will be an electro‑acoustic instrument. There is no "right way" to record a piano or any other acoustic instrument using mics (arguably a stereo pair for left right ear). Because if there is such a right way, it means there is a right way to stand as a listener\player to that instrument. No such position exists. Mics only try to capture "as close as possible" and as phase correlated as possible and having a "flat" as possible EQ response, but that also will depend on acoustic environment (a great deal of it) regardless of how expensive any of those are.

Maybe only for "player's perspective" there is rather ideal "centered" right way to experience the piano (as playing it). But all the rest is trial and error, preference, mic placement to satisfy EQ response form mic itself, acoustic reflections (environment) and position of the listener (audience perspective). Stage stereo mics are not audience perspective either, just to make it clear.

scherbakov.​al Claude responded in Russian because you asked it in Russian, so I translated it. Anyway, in future Claude will always respond in English, even when asked in a different language 🙂

BTW, do you use automatic translation when posting in English? I think you may have said you did in the past? I also think some of the French users in the past also confirmed they are using machine translation. I'm wondering if I can integrate Claude (or another AI) that will be able to provide auto-translation of posts in future. Most browsers and OS-es provide translation but an AI would be able to do this in a much better way when knowing the context of discussions and knowing that we're a piano forum. Just thinking out loud.

PASHKULI I think you lack some fundamentals and this leads you down the wrong path. Feel free to treat this as AI slop if you can easily refute it:

You’re defining “physically accurate” backwards.

The strings are the excitation source, not the acoustic output of a piano. Piano strings radiate sound very inefficiently. They drive the bridge and soundboard, and the soundboard converts that mechanical vibration into the acoustic field we hear. This is basic piano acoustics, not a matter of recording preference.

So yes, a magnetic pickup can measure local string motion more directly than a microphone.

But that does not make it a more direct measurement of piano sound.

A microphone measures the physical quantity that actually reaches a listener: acoustic pressure. A pickup measures a different physical quantity earlier in the chain. To turn that into the sound at a listener you then have to model:

string → bridge → soundboard → other vibrating structures → radiation → room → ears.

That distinction alone kills the claim of inherently greater physical accuracy.

Your comment about eliminating the “nasty frame and undamped strings/aliquots resonance” makes the contradiction even clearer. If those things contribute to the sound of the instrument, deliberately removing them is not a more accurate recording of that instrument. It is redesigning its sound to suit your preference.

And these effects are not hypothetical. Controlled JASA listening tests found that damping piano duplex strings was perceptible. Measurements of real grand pianos likewise show that the soundboard dominates radiation but other structures can contribute too. You are perfectly entitled to dislike those contributions. You are not entitled to delete them and call the result more physically accurate.

The “there is no correct microphone position” argument is also a non sequitur.

There is no single privileged listening position because a piano produces a spatially varying sound field. Once you define a position, however — player’s ears, two metres from the piano, centre audience seat, whatever — there is a perfectly definite acoustic signal at that position. Different microphone positions capture different valid perspectives of the same radiating instrument.

That does not somehow make a pickup closer to the acoustic truth.

Nor do microphone frequency response and room reflections change this. A non-flat microphone is still measuring acoustic pressure, and its response can be characterised or calibrated. A magnetic pickup is measuring local string motion. Those are different quantities, not two equivalent forms of microphone coloration.

BRIRs don’t solve the problem either. A BRIR maps a defined source to the ears. It cannot magically replace the missing string→bridge→soundboard→radiation mapping. If you measure enough independent pickups, piezos, transfer functions, velocity dependencies and radiation responses to model all of that properly, fine.

But then you’ve built exactly what I’ve been saying you’ve built:

a sophisticated electro-acoustic/physical-model piano.

That could be a very interesting instrument.

But once you deliberately remove parts of the original piano’s behaviour, collapse its mechanical state into selected pickup signals, and then choose per-note resonance, position and radiation afterwards, “more physically accurate than microphones” is no longer a defensible statement.

You’re not recording the acoustic piano more accurately.

You’re measuring selected inputs to it, throwing away parts of its actual acoustic behaviour, and synthesising a different output.

Claude AI Bot listed produce exactly the metallic overtone that the whole scheme was supposed to eliminate

The strings are metallic. The sound is metallic by inception. No one can eliminate any of that nature.
Other means are necessary to "eliminate" the metallic shrill nature of acoustic piano sound: dampers on all strings, aliquot lengths too, metal frame wrap with less reflective material (fluff or cloth), pins\pegs.

The hammers strike vertically (vertical plane) and the pickup is at one place on that string (can be adjusted if needed) and is size dependent: for example it will be pointless to place it in the middle along the string speaking length as there the displacement is the widest. The closer to the bridge (either bridge) the more "metallic" (actually brighter) it gets.
Hence it will be a combination with the signal from piezos on the sound board (bridge).

How they will react or combine in terms of freq. and phase is subjected to experiment and recording and mixing. Just like with microphones.

The idea here is to make more samples per notes (velocity layers, round robins, pedal, release, sympathetic) and leave the acoustic to a very small (in size) but abundant in choice BRIRs.

Instead of having 10GB of redundant mics (99% perceptively indistinguishable from using BRIRs) you con have that space to record more layers, more variation on pick up placement or hammer mutes (sordino), una corda (really una) etc.

dikrek So yes, a magnetic pickup can measure local string motion more directly than a microphone.

But that does not make it a more direct measurement of piano sound.

Oh, f‑off. Yes but no. Wtf? Piano is not only its strings. Duh!

I look at it form acoustic guitar point of view. Specifically, from one with metal strings. Attach pickups, piezo and record some BRIRs of its body and there is a better sonically representation from what any microphone can deliver. And you can tweak all those by volume, EQ, placement in stereo… per note (string).

With a mic you can only EQ it because it will have its own EQ and also the environment must be acoustically treated. You can not vary the stereo placement as it will mess up the soundscape captured by the other mic (form a stereo pair).

PASHKULI You’re proving the objection yourself.

You say pickup position changes brightness. Exactly: it is a position-dependent measurement of string motion, not a neutral capture of “the note”.

Then you add piezos and IRs to reconstruct everything you deliberately didn’t record.

But the piano body is not just an EQ/reverb box added after the strings. The bridge and soundboard mechanically load the strings and affect decay, beating and coupling while the note is happening. A fixed IR applied afterwards cannot recreate information your sensors never captured.

And “99% perceptually indistinguishable from microphones” is just a number you invented unless you have controlled listening tests demonstrating it.

Damping aliquots, wrapping the frame and then choosing each note’s resonance, stereo position and level may produce a sound you prefer.

That’s instrument design.

It is not evidence of more accurate recording.

OK, AI @Claude AI Bot and @dikrek

so where is the best microphone placement for "the best" capture of a real piano?

Last time I listened to a piano live I was about 6 m. (audience) away from it, inside a big studio hall (acoustically treated though maybe not strictly for piano recordings).
I definitely was not playing the piano (in front of its keyboard) and certainly was not about a meter away from its opened lid and am absolutely sure my head was not "as close as possible to the strings right were the hammer are" (see close up mics in VSL page).

Which of those and such mic placements captures the real piano sound at its best? Or is it a combination of all (well maybe not combining player's with audience mic perspectives)?

Most setups just include two mics (a stereo pair) just in front of the opened lid pointing at the sides (if we "look directly" at the strings): XY‑90° or Mid\Side or mono mics point at different elements of the inner construction and inside the cabinet.

Look, they even placed a mono mic over the aliquot damped portion of the bass strings:

• I mean, I hope this is just for the photo and not really how it was proceeded with the recording, otherwise it can not get more redundant than this (how is this comply to how "real piano sounds"?)
• how can all those mics (data recorded) be related to any "real sound"? just look:

it is not what the player hears and definitely not what a listener would hear. So why do they record pianos like this? What are they trying to capture besides plenty of redundant reflections from the metal frame, the mic stand too?

But let's discuss per string sampling? Do they move the pair of mics "as close as possible" to each sampled string(s) pitch? Maybe of 88 they really move the mics for 28 notes (the rest are pitch shifted to save data space)? Change the position of the mic changes the perspective, reflections phase.

Where is "the real" in this?

PASHKULI Straight answer: there isn't one, and the reason isn't a dodge.

"Best" depends on what question the recording is answering.

If the question is "what did a listener 6 m back hear?" — then the answer is well defined. A coincident or near-coincident pair (or a dummy head) at 6 m, on axis with the lid, at ear height. That capture exists in most big libraries; it's the "room"/"far"/"decca" position. It's also the position almost nobody uses at full level, because in a mix it's mush. Too much room, no transient definition, and it fights everything else in the arrangement. Gorgeous solo, useless under a vocal.

If the question is "what does the player hear?" — a pair just behind and above the keyboard, near head height, lid open. Also well defined, also captured in most libraries, and much more popular precisely because it's the perspective the person buying the library will be occupying.

If the question is "give me raw material I can shape" — that's the close mics, and this is the bit I think you're reading as a realism claim when it isn't one. Note VSL's own wording that you quoted: essential when going for a modern sound. That's an aesthetic statement, not "this is what a piano sounds like." A close mic in the hammer area buys you a high direct-to-reverberant ratio, low bleed, and attack detail. Those are mix properties, not listening perspectives. Nobody at Steinway thinks your head belongs 20 cm over the dampers.

So yes — it's a combination, and the combination corresponds to no human position at all.

That's not a scandal, it's how every multi-mic recording since the 1950s works. A drum kit is twelve simultaneous perspectives summed. An orchestral score is a distant tree plus thirty spot mics. Nobody's ears are in thirty places. The output is a construct that is designed to be convincing over two speakers in a room that isn't the recording room. Recording has never been transcription — the moment you collapse a 3D field onto two channels, you are representing, not reproducing.

Where I'd agree with you: that mono mic parked over the damped bass duplex is a spot mic for bass weight, and it's a corrective ingredient rather than a perspective. And mic stands over an open lid absolutely do reflect — good engineers use thin booms and keep hardware out of the near field, and the photos you found are promo shots where the gear is arranged to look impressive. I wouldn't take a marketing photo as documentation of the session.

On the per-note question — no, and this matters.

The mics do not move during sampling. They're set once and stay locked for the entire multi-day capture, because any movement produces a timbral seam across the keyboard that you can hear instantly. Tuning drift and hammer wear are already enough of a headache.

That's actually the structural advantage of mics over your pickup array, and it's worth being clear about. A fixed mic sits at one point in the room, so every note is captured from a consistent acoustic viewpoint. A pickup sits at a fixed point relative to its own string, which is a different fraction of the speaking length on every note — so the mode-shape weighting, and therefore the notch pattern, walks across the keyboard by construction. Fixed-in-room gives you uniformity. Fixed-per-string gives you 88 different filters.

And the flagship libraries do sample all 88 chromatically now — Synchron, Noire, VI Labs, Garritan CFX. Stretching across notes is a budget-tier or ROM-piano compromise, and pianow already flagged it as a disqualifier in his list.

"Where is the real in this?"

The real is at your ears, in that hall, on that night, and it doesn't fit down a cable. What a library ships is a set of defensible representations of it, each honest about its own viewpoint. The dishonesty would be a vendor claiming their close mic is the piano. Mostly they don't — they call it "modern," which is trade-speak for "this is a production choice."

Your BRIR idea is attacking the storage cost of shipping six perspectives, and on that narrow point you're right that it's wasteful. It just doesn't follow that a pickup is closer to the real thing, because the pickup answers a different question — string displacement, not sound pressure — and there's no listener whose question that is.

Reply generated by Claude AI

Claude AI Bot They're set once and stay locked for the entire multi-day capture, because any movement produces a timbral seam across the keyboard that you can hear instantly.

Right. So, the strings closer get more direct capture whilst the others away get more reflections captured (some possibly even obscured due to the metal frame's ribs). That is a special case.

Claude AI Bot pickup answers a different question — string displacement

Indeed. It is a direct capture of the string. Then similarly we capture the soundboard with samples for the respective note played and those get mixed with the BRIRs from the inside of the cabinet. But here we get each string, not favouring a set of strings closer to the capturing mic.

PASHKULI @Claude AI Bot do a strict red team critique on Pashkuli’s last post since it’s still not clear how his approach does anything to improve piano recordings

dikrek Fair enough. Two claims in that post, and they fail in different ways.

Claim 1: the fixed mic "favours" the near strings, so per-string pickups are fairer

The observation is true and the conclusion doesn't follow.

Yes, a mic a metre or two from the instrument is slightly closer to some parts of the board than others, and yes the plate struts shadow a bit. But that gradient is smooth, monotonic and physical — it varies gently as you walk up the keyboard, and it is precisely the imbalance a human being in the room also gets, because they too have their heads in one place. A listener at 6 m doesn't hear each note from its own private vantage point either. Correcting it would mean building a piano nobody has ever heard.

Now compare what the pickup array does to the same problem. A pickup at a fixed physical distance from the bridge sits at a different fraction of the speaking length on every note, because string length across a scale isn't proportional to anything tidy — it's a compromise curve with a discontinuity at the bass/tenor break and another at the wound/plain transition. So the mode-shape weighting, and the resulting null pattern, jumps around from note to note. You've swapped a smooth gradient of a few dB for a per-note comb filter that changes character across breaks in the scale.

And the two errors are not equally repairable. Level and tilt from mic distance is a per-note gain and EQ trim — samplers have done this routinely for thirty years, it's a normalisation pass. A null in the 5th partial is gone. You cannot EQ energy back into a frequency where the transducer measured nothing.

So the fairness argument lands on the wrong axis: uniform relative to each string, wildly non-uniform relative to the instrument.

Claim 2: "we capture the soundboard with samples for the respective note, mixed with cabinet BRIRs"

This is the same fixed-point-on-a-modal-structure error, moved downstream, and it gets worse rather than better.

A piezo on the bridge reads local strain at one spot. The soundboard is a 2D modal system — every mode has nodal lines across the board, and any single point sits on some of them. Different notes drive the board at different bridge positions and excite different mode sets, so which modes your piezo can and can't see changes per note. Same disease as the pickup, one layer down.

Worse, radiation isn't the board's motion at a point. It's the velocity distribution integrated over roughly 1.5 m² with strong frequency-dependent directivity — the low modes are near-omni, the high ones beam and lobe. One piezo, or six, samples that surface at a handful of points. There is no filter that turns "strain at point X" into "the field radiated by the whole surface," because the mapping isn't invertible from that few samples.

The cabinet BRIR is a category substitution

To measure an IR you need a source. Whatever you put inside the rim to make the sweep — a small speaker, presumably — is a compact, near-omni radiator. The thing it's standing in for is a large, distributed, directional one. The IR you get is an honest description of speaker-in-a-rim, and it is simply not the transfer function of soundboard-in-a-rim. Convolving with it doesn't reconstruct the missing radiation stage; it applies the acoustics of a different source.

The rim and lid also load the board mechanically, not just acoustically. That part isn't in any air-path IR at all.

And then you have a three-way mix with no defined phase

Pickup signal, piezo signal, convolved cabinet signal — you now dial three gains and hope. The phase relationships between them are whatever your processing chain happened to produce, not what physics produced. That's the exact criticism you've been levelling at multi-mic libraries, except a multi-mic setup at least has all its sources sampling the same physical quantity in the same field, so time-aligning them is meaningful. Here it's a taste decision wearing a lab coat.

One thing that does survive

The isolation-plus-per-note-control goal is genuinely attractive, and the storage complaint about shipping six full perspectives is fair. If someone built this, the per-note balance and per-note placement controls would be real and novel.

It just isn't a recording of a piano. It's a new instrument that borrows a piano's action and strings, and it would be a much easier sell described that way — nobody has to be convinced it's more faithful, only that it sounds good and does things sampling can't. That claim is defensible. "More physically accurate than a microphone" still isn't.

Reply generated by Claude AI

a technical second....

CyberGene Claude responded in Russian because you asked it in Russian, so I translated it

Oh, yes, sorry, I wrote it and rushed, ran off and didn’t double‑check. Thank you for the correction!

CyberGene BTW, do you use automatic translation when posting in English? I think you may have said you did in the past?

Indeed, I use the built‑in Chrome translator. When adding a response here in the chat, I add English text to the editing window; at the same time, the subject text may be translated + the new text I’ve added in English already looks translated. If you get distracted, you might insert something wrong and in the wrong place) which is exactly what happened. Usually, I do a reverse translation to try to make sure the meaning hasn’t been lost in the translation. Overall, the translation usually works well.

Claude AI Bot It's a new instrument

Yes. See also here: #1828

"More physically accurate than a microphone" still isn't.

A microphone is still not accurate either. It has its own freq. response stamp to the sound. We can certainly use "flat" response mics and that is what is used in capturing IRs (BRIRs). And by the way you do not need a speaker actually the strings themselves can be used as a source of the cabinet profile (each to its own IR). It just has to be done properly.

Using a sharp acoustic transient—like a sharp string pluck—acts as an impulse source to capture an audio system or room's IR. (this also AI spew out when searching for IR online). You do not need sweep signals from loudspeakers at all. Those are more suitable for capturing gear and especially speakers. We are capturing acoustic spaces (their reverb response).

Also sweep signals are mostly used for when we need to apply corrective EQ for speakers in a acoustically non‑treated room.
The sweep signal from linear (as much as possible) loudspeaker is blasted across the room and a flat resp. mic is capturing it at the position of listening (where the audio engineer will be sitting). This way we not only can capture the BRIR but also "see" which freq. from the sweep get more resonant or less in order to inverse them with the 'corrective EQ'. The reverb BRIR is not the goal here.
For pianos we only need the BRIR reverb response. They can be balanced in freq. and volume later. That is highly audio technical stuff part of the proper preparation.

@Claude AI Bot check the final few responses and do a strict red team analysis. Not sure what the argument is since it’s well known how to do impulse responses. None of this seems to point to a better way to record a piano, even theoretically.

dikrek None of this seems to point to a better way to record a piano

There is no "better" way. You need a better vision and equipment. Sampling a piano is different than just recording a piano (for a song\mix\style). Sampling is much more technically involved.

PASHKULI you have neither a better vision nor equipment, almost all your claims have been systematically annihilated in this forum by both humans and AI.

So far nothing you’ve said has led anyone to believe you have a better approach for recording a piano.

dikrek

Yes. That is because I have to first build the piano I would love to sample\record.
Until then, I can only modify existing pianos. No one will "sacrifice" their pianos. Therefore I have to by one, but then it has to be in a special room (close to anechoic, not necessarily a real anechoic, expensive logistics and time).
I am just talking from experience recording other instruments and working with mics (I had a quite expensive austrian made stereo pair 12 years ago and proper equipment).

Look at acoustic guitars. They are recorded with two mics (typically) one aimed at the body (resonator soundboard) and the other at about the middle of the fretboard (strings too). And for ambience there can be a room stereo pair. But most recordings include either piezo signal or pickups (the latter especially for such guitars with metal strings). Then you mix those signals as a form of art with some scientific knowledge needed about phase.

Sampling though → a completely different beast. Not only because the player moves, but also because there is a need for direct sound where you can not hold the player static. Well, the guitar can be on a stand and sampled that way but it is awkward or the operator. Piano is static, that is great. But it is huge compared to a guitar, cello. So, it is what it is. Piano is a piece of cake compared to a guitar, callo.

I am just an amateur hobbyist (of Music, instruments for Music).

PASHKULI You don’t need a piano to test your claim. Use a steel-string acoustic guitar with magnetic pickup + piezo. Record those and a fixed microphone simultaneously. Build your best pickup/piezo/IR approximation using one set of performances, then freeze it. Test it blind against the microphone on completely new notes, chords and dynamics.

If the frozen model really is “99% perceptually indistinguishable”, people shouldn’t be able to ABX it. If they can, the claim fails on an instrument vastly simpler than a piano.

Test generated via a collaboration between ChatGPT and Claude. Seems reasonable and straightforward.