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

You call for an end to a petty discussion (as did I). Then you try start a new one with a "report to the mods" threat?? Amusing. (Perhaps that was your intent? Or perhaps you don't know Godwin's Law?)

Don't try to move the discussion away from the issue by accusing me of 'taking part'. If you were really concerned about it you would have messaged me instead of posting. (Ironically I'm now forced to do the same)

Reminding people about being apropriate is not the same as partaking in the offense so don't try and make it about that.

Also Godwin's law is an observation about what will eventually happen in a lively discussion about a contested topic not something you can invoke on purpose.

So once again, please try to stay on topic and please don't try to derail the thread and please use PMs to direct any additional concerns about that to me.

Thanks

PM is no fun. It's just like FaceBook - having private conversations in public is where we are in today's society. Since this is a public blog and no rules are being broken, I say let the steel cage match continue.

No offense but it appears that this thread is now about bravado and rehashing.

My opinion on sample vs. modeling is that it appears a hybrid will be the best compromise for many years.

....or the Mrs Gamp episodes in Martin Chuzzlewit. Or even the part where Krook explodes in the middle of Bleak House.


ps - really agree with everything 36251 has just said. And this has, for me, been a very enlightening thread on modelling design & development.

As a Roland fan myself, I must admit I have also seen the (IIRC, official) Roland video of Scott Tibbs claiming a Bosendorfer as an attempted voice on the V piano.
Thinking about it they are not very likely to use a Yamaha, I suppose.

But even so, there is actually no definitive sound of a piano. Every Bosendorfer will sound different and as they age the differences will magnify.

Therefore, (again, in my opinion), there is not, nor ever will there be a definitive, end of story, ultimate, finished and forever piano tone from any manufacturer of acoustic pianos that will be the sample/model for all time.
Due to the analogue nature of acoustic pianos, even if there were, every subsequant article produced will be slightly different anyway.

For this simple reason an accurate emulation is sufficient to be considered passable, hence, Roland, Yamaha, Kawai, Korg, Nord, Casio... etc, etc, etc, all produce a passable voice for a piano tone.

Not only that, but (theoretically) that tone will never change.

This has an upside and a downside.... one, it will never improve with age, (you could take an arguably legitimate veiw that software upgrades could in fact improve the tone, contrary to my statement) two, they will never detiorate with age (you could take an arguably legitimate view that electronics components do age and deteriorate over time so my statement is not accurate).

Either way, the majority of top range digital pianos have reached a stage where they are acceptable, pretty obvious really, if they hadn't no one would buy them.

I am certainly not saying there will never be an improvment, but how many people really do not buy an electric piano because it sounds so different to an acoustic?

There must be some but my view is that they are the losers, as I believe accuracy of touch is more important than the tone produced anyway.
The greatest advantage of a digital piano is that when you discover something in the tone you don't like, you can switch the voice. You can't do that with an acoustic, although I did once put drawing pins in the hammers to try and make it sound like a harpsichord!

I have owned acoustics from 1971 to the present day, I owned a Roland RD 300s from 1991 to the present and have had a Roland RD 700NX for about 19 months.

The consistancy of the Rolands makes them far more enjoyable than the acoustics and the quality of feel and touch is vastly superior.

I have an RD 700NX too, and it's great for what it is.

But I will NEVER, meaning NEVER, willingly play it in preference to my acoustic. I'm glad to have the RD for when I need silent practice or to play whizbang games with electronics, but when I want the best piano playing experience available (which is almost always what I want), I play the acoustic.

I often get the impression that many DP users have only ever played one or two acoustic pianos (if at all) and base their idea of 'piano sound' (and 'piano action') around that, or maybe from one or two favorite Youtube videos. Bösendorfers can sound anything from mellow to strident, even within the same model, and the same goes for almost all other brands, except maybe for Blüthner, which I've never encountered a strident-sounding one. The age of the pianos play a part too, as well as how they've been voiced. Yamaha and Kawai tend to be more consistent.
And the acoustics of the room where you play the piano/DP (if not using headphones) play a big part too, which is partly why I always base my assessment of DPs using my own headphones - the other reason being the inadequacy of the amplification and speaker system in many DPs.

A DP that doesn't feel and respond anything like the real thing is, to me, pointless - no matter how good its inherent sound, or how 'authentic' its action. As many people have discovered, it's all too easy to get bored with a new DP after just a short period of time, and I believe this is partly to do with its predictable sound picture. Which is where a modeled DP has the advantage: it can give you an ugly sound or a beautiful sound depending on how you want it, and how you play it. Whether the sound is infinitely variable or not (according to some posters here, apparently it's no more variable than sampled DPs), it feels like it is, and I've never got tired of playing and practising on my V-Piano, because it responds to my touch - good or bad - just like an acoustic.

Last weekend, I made some recordings on a Yamaha CF6 grand for a piano showroom website (and possibly a magazine too). Even though it wasn't a CFX, it sounded very powerful and loud in that showroom, in fact, louder than the CFX I played a few weeks ago (in a much bigger hall). I had to do a few takes before I adjusted to it. It was all to do with its voicing and the acoustics.

You are making my point for me. Pianoteq works fine on a Core 2 Duo, and doesn't come close to exhausting the computing power of standard state-of-the-art 12-core desktop computers, let alone take advantage of custom hardware implementations with DSP processors and FPGAs that could be used in standalone DPs. My experience is with implementing 3-dimensional spatio-temporal video signal processing, software algorithm design, and custom DSP hardware design using FPGAs.

Pianoteq is in no way restricted by the computing power of today's desktop computers, and built-to-purpose standalone DP's in the V-Piano grand price range could have many, many times more DSP horsepower for computing far more complex DE models and algorithms than desktop computers. Computational power is in no way a limitation for producing better modeling performance when today's "best" implementations can run on Core 2 Duo processors. What is needed is more realistic modeling.

I think the term compute-intensive, applied to a program, is meant to say that it mostly performs computation, and that I/O is a lesser factor.

It need not mean that the program stretches the capability of any typical computer.

I run Pianoteq (or rather, I used to) on a beater laptop, a dual-core CPU from 2006, 1.7 GHz. Pianoteq ran with no problem at all. Today you'd be hard-pressed to find a computer that lame (except maybe for the low-end netbooks). Pianoteq may be compute-intensive, but that's only in comparison its I/O-paucity. Unlike samplers, it does not spend its time reading digital audio files.

Doing a custom DSP and FPGA Setup would require a much higher volume production run than you'd get for DPs

In sub 20,000k runs even a standard analog devices blackfin DSP is $20 and most silicon manufacturers won't even talk to you (call us when you want a million +) a capable sharc or tigersharc will be above $50 or even $100.

Thats for a single DSP core.

a capable xilinx FPGA core would also cost you roughly $50 to $100

This doesn't include any of the parts that actually make up the PCB of the piano and it probably still wouldn't get you anywhere near the performance of the i5.

Such a custom made board would probably be around $500 for the complete bill of materials or twice that if you want anything close to PC specs.

So you'd have to retail the finished PCB alone for three times that to come up with even a slight profit margin.

Then you'd still have no knobs and buttons, no enclosure, no action, no display and no software.

So you end up with a device that may be close to the performance necessary to run a pianoteq model at about the price of a V-grand

DSP setups are not about max performance it's about comparable performance for less, less heat, less money, less overhead less pcb space.

It can be cheaper if we're talking about decent economies of scale but you'd still only be able to match the core2duo's performance for a higher price but less complexity and less power consumption, smaller form factor. But also less memory bandwidth and less overall memory.

Also diffentials are hard even for floating point DSPs and might even be more taxing than video filtering/analysis and en-/ decoding. DCTs and FFTs are usually builtins, complex differentual equations are less common though so not that readily supported. (I assume your algorithm does encoding and video analysis/processing in the frequency and time/ spacial dimensions)

For me a current quad core as recommended platform falls under 'computationally expensive' especially when the result is so underwhelming. To match the power of your state of the art 12 core would be prohibitively expensive to do with a custom DSP setup. Look at abject failures like the alpha that will sell for a MSRP of $30000 if it ever gets made.

this also assumes that the modelling algorithms are as easily parallelizable as picture/video analysis/processing algos and can use all of the cores.

It's not only about what's available it's also about cost. Nobody will build a DP that could run current state of the art models if the price would be too high, that's because nobody would buy it.

Even devices like the V-Grand are enthusiast instruments for a very small demo.

That doesn't mean it also could be about the quality of the models.

The $1000 question is though, are the models the way they are because of their lack of maturity (possible and probable) or are the models that way because it's the best you can do on current hardware?

Recomending a decent quadcore but only 256 megs of ram smells like the CPU is the bottleneck here ( I'd guess it's an NP hard problem or at least exponential time)

My opinion is that it's both column A and column B.

Just for reference: Your state of the art 12 core would be a 2010 amd opteron at 2.5 GHz per Core and a MSRP of 1200 dollars.

Buying a computer at retil prices ith tht kind of performance will be in the 5000 to 10000 dollar ballpark.

A similar setup with two Core I7 6 core would set you back 4000 dollars just for the CPUs. The whole system also consumes about 400 to 500 watts of power.

Scaling a custom DSP system to that performance would probably be similarly expensive and would require probably more than 12 DSP cores. (more power per MHz but taps out at 400 MHz instead of 3 GHz) You'd also probably require a fan setup and a huge PSU

When you look at the kind of hardware that's used in current DPs and workstations I don't think that this would be in any way feasible.

The kind of computing hardware used in current standalone DPs is archaic, but that is besides the point.

I said that "today's modeled pianos don't come close to using the CPU power available in today's most powerful computers". That is a simple fact. Pianoteq runs on a Core 2 Duo. That isn't even close to the computing power available in today's 12-core standard desktop computers. So Pianoteq doesn't come close to using the CPU power available in today's most powerful computers. Hence, my point is proven.

Your comments about price and power of desktop computers is totally irrelevant to my point.

FPGA's are ideal for small volume production (I use them for products that sell a few hundred units lifetime), so I don't know what in the world you are talking about. The cost of custom silicon is prohibitive for small volume, which is why we use FPGA's for small volume. That and the ability to upgrade algorithms and performance in the field. But this is not a debate about manufacturing a DP at a particular price point, or with a particular margin. It is not a debate about whether a desktop computer costs $4K, $6K, or $10K.

You are hung up on cost and that is simply irrelevant to my point that the state-of-the-art in modeling is limited by algorithms, not computing power. Show me an implementation that produces much better sound quality that ONLY runs on a super computer. That would make your claim that much better algorithms are limited by computing power. I haven't heard of any such implementation. On the other hand, today's current algorithms (Pianoteq and V-Piano) run on computing solutions far inferior to what is available in off the shelf desktop computing, or that could be implemented using custom computing solutions.

FWIW, I do clearly recall a Pianoteq person saying that they wouldn't know what to do with extra CPU power, which does support what Macy is saying. It was in the Pianoteq forum on their web site. (sorry no link at the moment)

Greg.

Yes, a first prototype that truly is authentic in regards to the real sounds of a concert grand, with no reservations in cost so that R&D can perfect it. Someone would need to make a substantial donation so that the work could be undertaken.

Just one sound needs to be "perfected" with modeling (perhaps a Steinway D, or, Kawai EX) so that the other modeled parameters would complement the sounds. Also, a purely modeled instrument needs the best projection of its sounds with a revolutionary speaker system equal to the task.

Questions:

1) So, the question is, how much money would realistically be needed to complete the project?

2) Exactly what would this prototype look like... as for cabinet, speakers, key action, etc?

FPGAs are a sensible choice if you need custom silicon but can't afford to order an ASIC and you don't have to concern yourself too much with PCB size and power dissipation.

It's an alternative if a custom ASIC would not be financially feasible, it's more expensive though than the custom IC would be at volume and its also more expensive than a component off the shelf would be if you don't need the custom hardware.

Low-volume production at the gross margins of the instrument makers will probably use as many off the shelf components as possible.

Look we can argue all day about technology when all you do is focus on a single aspect of the product.

I'm 'hung up on cost' because in the real world people have to buy stuff with money they have to earn first.

If all I focus on is performance I could build a killer rig with the processing power of a small super computer that probably wouldn't be able to be exported from the US due to its classification as military hardware. It would also cost a lot of dough.

Then you end up with something like the Alpha piano at $30,000 or a V-Grand or Avant-Grand at $10,000 to $20,000 that still don't sound great.

At that price I could simply go and buy a real baby grand instead.

If we're talking about DPs here then there are additional concerns.

It should fit into an enclosure roughly the size of current DPs
It should be able to work without a fan or at least with minimal additional noise
It should be able to do at least the same things competing products do
It shouldn't break the bank.

The reason for that is that even in a small volume business sector that's largely populated by enthusiasts, instruments in the $4000 and above category are considered to be too expensive and somewhat aspirational (if I win the lotery...).

Kawai probably ships a great deal more $600 Clavinovas than MP10 in a business where even enthusiasts consider a Nord Stage 2 or Kronos X to be a luxury item.

The Kronos by the way uses a dual core Intel Atom 1.8 that isn't supplemented by any custom DSP hardware as its processor supplemented by 2 Gigs of Ram and 64 Gig of SSD storage which would cost you probably less than $300 retail yet it's still $3999.

The newly announced Prophet 12 uses 6 Analog Devices Sharc DSPs 'just' for the analog modelling of the synth (> $400 gross prices at 1000k) at the low price of $2999 and the forum buzz is along the lines of 'I would buy one if I could afford it'.

So cost and performance per dollar spent is an issue: Not from a technical standpoint but to make it a viable business case.

That doesn't mean that you're wrong though.

I just wonder if modeling is still as much in its infant stages as you claim or if they are simply limited by what they can do with their hardware.

If you look at computer graphics, rendering and CGI or video processing, tasks that are probably similarly power hungry then they are all severely limited by the available hardware or budget or my business sector where companies spend insane amounts of money just for a little more performance to do physics modelling with FEM a little better.

I'd be great to get some of the pianoteq foks to talk about their take in this issue.

Need to see modeling implemented so that a digital really sounds like the acoustic pianos that are being mentioned.

Roland's examples for "Steinway D" and "Bosendorfer" in the V-Piano are still well off the mark:

http://www.youtube.com/watch?v=7Ewq6NgYpxA

Does anyone hear a "Bosendorfer" in any of the "Vintage II" presets?

As I currently own a V-Piano ... me thinks not.

These type of projects are commonly done by university researchers. I've read a few papers on acoustic piano modeling written by university researchers, but I don't claim to really be knowledgeable on the subject. (I simply made the obvious observation that today's implementations don't take advantage of today's state-of-the-art computing power, off-the-shelf or custom. I don't know why that created such enthusiastic arguments when it is so obvious.) Anyway, universities and PHD students are a great place for this type of work to be done. Prototypes don't have to be in practical form factors - this type of research rarely ends up in commercial form at the university level, the point is modeling and algorithm development. Then private companies can take the results and commercialize it in practical, cost effective form factors.

I've been developing state-of-the-art products for about 40 years. In my current field of expertise (about the last 15 years or so) research (university and private) has always been far ahead of what could be commercially implemented - and still remains so today. We are compute bound and have always been, i.e. our most advanced algorithms can simply not be implemented in real time with the most advanced technologies available. However, we have continuously implemented ever increasing complexity algorithms over the last 15-20 years, implemented first with exotic custom hardware for cost-no-object defense markets, or cost-nearly-no-object professional markets, and then eventually implemented at more practical price points for broader markets.

What I haven't seen in piano modeling are similar research demonstrations that show that truly realistic piano models are known, or that subsequent algorithm requirements, if such models exist, are compute bound. i.e. before we can blame computer resources show that a solution exists, and then we can work on the computing issues. There are only a few commercial piano modeling implementations in existence and they are executed on what is very basic consumer computing platforms that don't require state-of-the-art processors, DSP's, or custom hardware development. That suggests that better (probably more complex) modeling solutions haven't been developed yet (else we would have demonstrations or products requiring state-of-the-art hardware) and the focus is still on developing modeling solutions, not beating computing limitations.

Since design digital piano (sampled and modeled) from a long time, I have read with a lot of interest this Forum.
Many aspects of the topic have been discussed and some of them quite deeply. I would like to replay at many of them, but I prefer to give my answer, based on my personal experience, to the initial question.
First of all, not all the piano are sampled, since on the market there are the modeled V-piano and Pianoteq, while starting from the next month, also the our Physis Piano will be available.
So: Why all pianos are sampled not modeled?
My answer is because :
• The sampling technology is cheaper than physical modeling. For example for the Physis Piano we use six DSP, while, typically, only one is necessary for sampled pianos.
• The sampling is a mature and widely diffused technology that is quite easy to understand and to implement, while the Physical modeling is really complex to comprehend and calculate. Just to give an idea a I shortly tell my experience: I started to study the piano physical modeling 20 years ago. At that time the advancement on the research and the computational power was not enough to reach an acceptable result to design a good modeled piano, that was why I abandoned the project. In any case this first effort was not useless since those studies were useful for us to “discover” and apply the string resonance to a digital sampled piano. Seven year ago, in cooperation, with 3 universities and the extraordinary team of my company, we started again to design a modeled piano. It has been a real extraordinary effort, but this time I think we got a good result.
• So on the basis of my experience I think that, right now, almost all the digital pianos use sampling because it is too hard and expensive to make a modeled piano.
In order to answer to the original question in some post there was a discussion on what it’s better : a real acoustic piano, or digital sampled piano, or a digital modeled piano ?
The answer is: it depends.
All digital pianos (sampled or modeled) are an imitation of “the real acoustic piano”, therefore they approximate, in some way, the reality. If we strictly interpret this sentence we could say that the reality is always better of whatever reconstruction, but, if we detail the answer we could say that:
• We have not to consider the “real acoustic piano”, but complete panorama of the acoustic pianos, starting from the worse vertical Chinese to the best grand Steinway. In this case it is easy to find that many digitals are better than some acoustics under several points of view.
• Digitals pianos have well known additional features that acoustic don’t have : tuning not needed, headphone play, more sounds, sequencing and so on.
• Digitals pianos are cheaper than acoustics.
In synthesis there many reason to prefer a digital piano to an acoustic.
But if you prefer a digital piano what it is better: the sampled or the modeled ?
The answer is still : it depends.
• If you want cheaper piano: sampled.
• If you want adjust the piano sound at your taste or your played music: modeled.
• If you want a sound very close to the real piano: sampled, seem to be the right answer, since a recording is supposed to be always more accurate representation of the reality in comparison to a complex reconstruction, even if this is based on a very accurate modeling. This answer could be right (it depends also how you make the recording and reproduction) if you consider just a single note played at specific key velocity, but it is wrong if you consider many notes played at different velocity, i.e. if you play a piano. In fact in this case the interactions between notes, i.e. the sympathetic resonance (known also as string or damper resonance) become a very important phenomena that make the whole sound really true. In the modeled piano this effect is accurately reproduced, while on some the sampled piano it is not reproduced, or, in some others, is reproduced in a rough way. Moreover all the sampled pianos suffer of sample granularity, i.e. they use the same sample inside a note and velocity interval. These intervals, depending on the avilable memory, are wide in cheap pianos, while are small in a computer pianos, but always exit! In the modeled piano this problem do not exit at all, since is the modeled piano is practically continuous. There are other advantages on modeled piano that allow you to truly reproduce the note rebound or release. http://www.pianoworld.com/forum/images/icons/default/lightbulb.gif

Tough to gig with, though. ;-) Really, there are a whole lot of reasons people buy digital pianos, it's not always because they're cheaper than the real thing.


I suspect the former. Especially since there is no available piano modeling software that requires even the best hardware that is commonly available today. Though who knows what prototype next gen V Piano Roland may have in the labs, waiting for the required hardware to get more affordable before they can make a product out of it.


One difference between acoustic modeling for piano and CGI is that CGI can take as long as it needs to render a frame, and eventually, you get something useful out of whatever your code was capable of. The piano model is absolutely useless if it can't work in real time. So right from the start, whereas there was a reason to develop CGI that would do what you needed practically no matter how long it took (i.e. it would still give you a marketable product), there is no reason to develop a piano model that is too far beyond what current technology can do in real time.

PIxar started as a hardware company, and they almost went out of business. Creative content saved them. With the potential benefits of CGI clear, it was worth the ungodly amount of money that has been poured into CGI over the last 25 years or so (despite which, it's still not capable of creating everything in a way that is indistinguishable from the real thing). The amount of money to be made by modeling the sound of a piano is rather limited by comparison. For any company to invest tons of money into developing new technology, there has to be a sense of how they think they will be able to make their money back.

(edit: though Macy makes good points about university research as well.)