Piano World Home Page Forums Our Most Popular Forums Digital Pianos - Electronic Pianos - Synths & Keyboards Why all pianos are sampled not modeled?

That's EXACTLY what I mean.



This would be entirely and undeniably true if we'd be talking about the real instrument.

I won't even disagree that modeling might sound better under certain circumstances.

I just sense some sort of confusion here about how digital instruments (regardless if they are modeled or sampled) actually work and instead I hear a lot of conjecture about how they're supposed to work.

High quality sampling doesn't simply "just sound like 8 notes all distinct from each other" there is some modeling and synthesis going on.

Most companies model sympathetic resonance for example or improve the sampling with synthesis and modeling to blend samples for different velocities together etc.

So high quality sampling that is helped by modeling and synthesis sounds better than you give it credit for.

While a pianist might be able to "produce more variations of dynamics" than there are samples, the digital piano with modeling will not be able to reproduce all of them.

It simply can't.

You play on a simulated key action. The sensors that measure the key travel and velocity convert that information into digital numbers of finite length and resolution (say 16 bit or 24 bit) so right there your theoretical limitless number of "variations of dynamics" is converted into no more than 65535 different levels (for 16 bit) per keystroke.

After that process the modeling algorithms use that digital information to calculate how a real piano would sound like. The quality of that process is determined by several factors.

- Is the modeling realistic and complete enough to recreate the sound of the real instrument within a quality threshold that makes the difference unnoticeable.

- Could my hardware even run such a model if it exists.

- is the fidelity and resolution of my digital system good enough to actually reproduce all I want it to reproduce.

What I object to quite simply is the notion by many people in this thread that modeling is inherently 'better' because only modeling is able to really recreate the sound of an acustic piano because they attribute some sort of magical properties to it that supposedly sampling won't ever be able to match.

Both methods are 'digital representations' that lose information.

One or the other might sound better or more real to you depending on the current state of the art of the competing technologies. I won't even argue about that.

If you factor out business decisions, feasibility and processing power however as some people seem to do to make their case for modeling then there is no inherent reason why one has to be better than the other. After both are good enough so that the human ear won't notice the difference then it's a matter of preference and not quality.

If you live in the real world and have to consider things like price, technical feasibility and your target audience however then there are differences.

So to get back to the question of the OP.

Most instruments today are sampled because it's cheaper, sampling is less taxing on the CPU and hardware so you can offer more features in a smaller package and the technology is more mature.

It's also much cheaper to simply spend the effort to record a piano instead of spending a huge budget on R&D for a good modeling algorithm especially when you can improve sampling with synthesis and modelling parts of the instruments.

Thirdly for most use cases (band context) the supposed higher fidelity is wasted due to environmental factors (fidelity limits of PA, amplifiers, sound reproduction properties of the hall you play in etc.)

And last but not least in contexts where that matters (concert halls) people most probably would use a real acustic.

So it's entirely a business decision by the companies.

For all practical intents and purposes: Yes.

In Theory: No.

The harmonic series of overtones of a base tone goes on to infinity, so in theory the n-th overtone of say a c' is still part of the sound for arbitrary n even if n -> infinity.

Since the sound of a piano is the complex combination of all sounds including all of their harmonics plus the interaction with the sound board, corpus and resonating elements, the number of different combinations is truly 'infinite'.

We simply agreed though, that after a certain point the additional 'information' is below the capabilities of our sensory organs to notice and becomes 'unnoticeable' by human ears so it doesn't mater if it is left out.

The lively debate about digital vs. anaolog shows however that those limits might be a little too arbitrarily defined

You could spend a fortune on a acoustic concert grand and get a piano you don't like, so why do many people pretend that acoustic always equals perfection, in the same way they pretend sampling is always inferior to modelling?

It very often isn't. The vast majority of digital pianos are perfectly passable replicas of acoustics.

I cannot agree with this:
I think most digital pianos sound piano-ish, but are not perfectly passable replica. Not even close.

I'll go as far as saying that they're usable. But I want something MUCH better.

Piano sampling software is a big step toward improving that, and I'm satisfied. Some of them are perfectly passable.

But I can't say that about the modelers (Pianoteq). Their Bluthner is something of a breakthrough. If they can take one more step up, I might find it perfectly passable. (Not perfect. But passable.)

No. The vast majority of digital pianos are somewhat close to an acoustic piano in terms of sound and action so as to make for a somewhat satisfying pianistic experience. Calling DPs replicas of acoustics is a massive overstatement.

I own both an acoustic piano and several digital pianos and am glad to have and use them all for different purposes. I'm not on either side of the "debate," since for me there is no debate.

I do think, though, that we'll be able to say convincingly that a digital piano is a passable replica of an acoustic piano when we're able to run blind listening tests in which experienced pianists cannot distinguish between the sound of a digital and that of an acoustic.

With the qualification that I'm largely innocent of software virtual pianos (so far), I'm not sure we've reached that point yet.

Well, that hasn't happened yet though, has it? All current DPs fall short in the sound department - some are better than others. Decays are short, resonance is below par. That's just as a listener, but it's when you play on one you realise how lacking in life DPs are. There is a certain sterility and deadness to the sound. The tone colour isn't there either. I find it amusing when people can't tell the difference between a real piano and a DP - it calls into question their level of musicianship more than proving the DPs are "perfectly passable replicas" of acoustic pianos. There's a fair way to go before that can legitimately be said.

DPs do certain jobs well, but they haven't cracked the magic of a real piano yet - there's considerable work to be done first. I'd imagine the new Kawai VPC Or AG + a top software piano would be the closest thing we have so far.

Why is it so forbidden to criticise DPs anyway? Why shouldn't we have high standards? Are we not allowed to be discriminating when it comes to music?

My girlfriend use to say digital piano is like a dildo compared to real thing,
If you get her drift... It's actually pretty good analogy. Easy in use, portable and you can use headphones but other then that...

That comment is just RIPE for follow-up.
Wait ... too easy. Skip it.

Haha, gold! (Love the smoking icon at the end too)

With 128 levels limit imposed by MIDI I don't think this is a factor at all.



Well said. If there as to be an advantage between modeled and sampled is that one: interactions between notes. You can sample 88 notes at 128 levels, but if the sound of a previous note alters in any significant way the attack of a new one (and I think it probably does), then it will be really difficult and expensive to sample all the possibilities. And let's not even talk about the damper pedal at every possible level.

So, forget everything you said about steep and steeples because IMO the difference is not there when you just play a single note.

Carlos

Sampling and modeling are so diametrically different in concept to each other that they seem to appeal to different people, who may be looking for different things in a DP. Regardless of what or how 'similar' they really are after you take away the initial source of their sounds (as implied by previous posts), the fact remains that they do respond and behave differently when you actually play them. And dewster's tests in his DPBSD project also show up the marked differences between sampled and modeled DPs.

All my classical pianist friends, when they try out my V-Piano, say things along the lines of 'it's the only digital that feels and responds like a real piano'; and that if they had to exchange their acoustics for a digital, it's the only one they would consider. On the other hand, my two pop and jazz-playing acquaintances weren't at all impressed (and they are the ones who play regularly on DPs, and own several themselves) - partly because, as someone here once said, 'it's a one-trick pony', and it's 'poor value for money'.

Well, you pays your money and you makes your choice......

I agree. Within the 126 levels supported by MIDI (0 = silence, 1 would be key down and no sound, for proper piano behavior implementation), I don't see any inherent advantage to modeling over sampling for duplicating the piano sound from a given point in space, short of being able to do it with less memory. You could model different mic placements for different variations, rather than having to sample each of those mic placements, or different lid heights, or how perfectly in tune each of the 200+ strings are, or how worn the hammers and felts are, etc.... but these things give you more piano sounds, not necessarily any single better one.

And contrary to what someone said, modeling has no inherent benefit in creating longer decays. It can arguably create a longer, more natural decay if you're working on a system with less memory, but that's due to a limitation of hardware, not a limitation of sampling.

And you really don't have to sample more than 126 levels even for this theoretical perfectly sampled piano. The dynamic range of an acoustic piano (the difference between its quietest note and loudest) is probably under 63 decibels, so sampling at 126 levels would permit the samples at different velocities to be within a half decibel of each other, which is just about the limit of the smallest difference anyone could perceive. (Though getting the velocity response right is another issue, and I think this might be where being able to generate more than 127 values and map them accordingly could be valuable.)

Of course, whether you have electronics/amplification/speakers that can produce the entire 63 dB range of an acoustic piano at real-life levels and without adding distortion of its own is another problem altogether.

Anyway, it seems to me, in terms of a single accurate piano sound, the advantage of modeling only comes into play when replicating the probably infinite possible interactions between multiple notes. The sound of striking and holding middle C by itself could be captured in 126 30-second (or whatever) samples. But without the damper pedal, the sound of that strike will vary with whether you are holding down 1, 2, 3, or more other notes when you strike it, and specifically which notes they are, and possible how loudly those previous notes had been struck in the first place; and if the pedal is down, it might change depending on how loudly each other string is sounding, which would be affected by the sequence and velocity of each and every other note you had struck since depressing the pedal. That is, the "resonances" which the OP implied were unimportant are, in a sense, the only real inherent sonic benefit to modeling in the first place, as I see it.

Which is not to say that anyone has successfully modeled all those string interactions. Just that I could see modeling being a better solution there than infinite sampling!

Infinity is Infinity. Piano is an 'analog' instrument so it doesn't matter how you replicate it digitally. The closer you try to get to the real instrument the more you have to do.

This approaches infinity for both techniques modeling and sampling

Modeling all of the interactions between the string requires you to calculate all of the interactions between all of the strings that are involved in the sound.

This is the complex cross product of all interactions which is a mathematical problem whose complexity increases exponentially with the number of strings.

Each string you add has an influence on all of the other elements already present in the sound so each string you have to model adds a number of calculations to the model.

One for itself, one for the effect each of the other elements have on this string and one for the effect this string has on each of the other ones.

Modeling vs. sampling is essentially just a trade off between memory consumption vs. computational complexity (CPU consumption)

You either spend lots of memory for samples so that you don't need a powerful CPU or you use a powerful CPU so that you don't need a lot of memory with modeling.

A good computational model might be easier to tweak though but sampled pianos also do some modeling to let you customize the sound or to add effects that are difficult to reproduce with samples alone (string resonance for example)

So we'll probably see more hybrid approaches in the furture that use quality sample sets as base and modeling to make it more life-like.

Right now fast storage and RAM is so much cheaper than CPU power that it's easier to simply throw gigabyte after gigabyte of samples at the propblem and supplement it with modeling than to implement a realistic model of a certain complexity

what are we talking about here? probably 99% of all music is being listened to on and through electronic/ digital media so whether it comes from acoustic or digital instrument it doesn't really matter because it'll be converted to digital signal anyway and nobody can tell what was the original source. So why bother?

It's not about recording it's about playing.

A lot of people would gladly own a real upright or grand piano for practice and playing. Alas for many people owning a piano is not practical or feasible.

You need the money to buy and maintain one, you need the space for the instrument, you might not be able to practice at certain times because of the noise. It's heavy and you can't move it to gigs/recitals.

So people go digital. That doesn't mean however that just because you bought a digital piano you don't want something that sounds and feels like a real piano.

Anybody who's played a real piano at a reasonable level can hear the difference. And sitting down playing it, it's a no brainer. Unless you are talking about pianos sitting in a band mix.

I think sampling is not going anywhere fast, but the resonance modelling needs a lot more work. It also seems to me that by the time they are able to model resonance well enough to cope with all the vast complexity of dozens of strings interacting with each other with the pedal down, that will be the time when they are probably also good enough to model all the notes too. Until then, samplers still sound decent, even though the resonance is a bit disappointing. I do think full modelling will eventually take over, but perhaps not for quite a while.

And not when there's still a complete monopoly even now, 4 years after the V-Piano was introduced, and the biggest and most successful DP manufacturer (Yamaha) still content to rest on its laurels, just tinkering around the edges.......

Progress is slow, this is not trivial stuff. Pianoteq is the main software based modeling piano, and they're not worlds ahead of where they were four years ago either, and unlike Roland, they don't have the overhead of having to design a whole hardware system to go with it. (And I think some people still prefer the Roland even today.) So many people here seem to think that, if they can conceive of it, an engineer should easily be able to do it!


Yamaha's SCM does add modeling to their samples (CP1, CP5, CP50). I wonder if we'll see anything new at NAMM.