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

Some people love them, some don't, that's what makes the world go 'round...


Could happen, as modeling gets better...


Nonsense.

I think there is a flaw in your logic, if I understand what you're saying. I'm not familiar with computer language, so I'll have to explain using my own simple terms.

It's the difference between stepped and stepless sound. The sampled DP has a set number of pre-stored sounds from which all its sounds are based i.e. processed to get the in-between sounds that are required between the pre-recorded 'steps'. It's like the difference between stepped auto-focusing for early autofocus point-and-shoot cameras (assuming you are old enough to have used them), and the stepless auto-focusing as used in SLR cameras of that time, and of course, universal today, where the lens has no pre-determined step but focuses exactly on the object you're pointing its sensor at. For those point-and-shoots, the more steps the better, but the focusing was never totally precise, because there was a finite number of steps. (My first such camera had just three steps - close-up, 'portrait distance' and infinity/landscape, and relied on depth-of-field to give acceptable sharpness.)

OK, in the sampled DP, the computer helps by processing the nearest 'step' to get an approximation of the sound required by the particular key strike, but it's still a processed step, and it shows in its response, which doesn't quite ring true, and therefore reminds you you're playing on an electronic instrument. Or, to put it another way, you play a key a certain way to get a sound of 5.347, but there are only 10 samples recorded for that note (i.e. 1 to 10). So the computer chooses the nearest, 5, and processes the sound to get maybe 5.3 which is close but not exactly what you're looking for. And you 'detect' that the extra 0.3 isn't quite from source, but added to it.

When you play one note on a sampled DP and repeat it several times exactly (as far as humanly possible) without pedal, you hear exactly the same sound. It's quite uncanny, and you just don't get this from an acoustic, because other factors come into play - the very slight differences between each keystrike get magnified by the resonances that are generated. And this is also what you get from the V-Piano because it too is able to pick up on the very slight differences, which subsequently get magnified by the random resonances. It is (or rather, you are) a human playing after all, not a machine, and the 'butterfly effect' takes over.

The first part of your description essentially criticizes the fact that there aren't many layers in a sampled piano. American D, for example, has only 20 or so. MIDI allows 128, so that is what you get with PianoTeq or (afaik) the V. Whether the difference between 20 timbres and 128 is meaningful could be debated. Of course, with layer blending, you essentially have 128 distinct timbres with a sampled piano as well (Ivory uses some kind of layer morphing/blending according to their marketing). Of course, there are also sampled pianos with a full 128 layers (Imperfect Samples Fazioli extreme) for example. In that case you have precisely the same number of timbres as a modeled piano. You don't see people dropping Ivory in favor of Imperfect Samples wholesale, though. In part that's because the extra granularity in layers doesn't make much difference. Of course, if you are comparing the V to a P155 with 4 layers, the difference in granularity is much larger.

I can't really argue with your point about the timbre of a note changing with repeats. That's probably right. I would classify it as one of those things that sampled pianos to date have not yet implemented as far as we have seen. Historically we have seen sampled pianos improve as quickly as modeled pianos (or some might say more quickly) so I suspect these types of nuances will appear in sampled pianos before modeled pianos satisfy really picky people.

But you are right, some people are more sensitive to limitations of sampled pianos and others, modeled. Thankfully it appears that both technologies are improving. Personally, I'd love to own a V. It would make a great controller for Galaxy.

So are you saying the world goes round???



Of course.

Except that a computer algorithm that 'models' a piano isn't 'stepless' either.

That a modeled piano is 'stepless' is your assumption, one that has no technical merit unfortunately. You see for an algorithm or model to be truly 'stepless' it would require a system of infinite resolution - a truly 'analogue' system.

Unfortunately we're squarely in the realm of digital systems, that take and process discrete, quantifiable values. These values have a finite resolution and there is an upper bound on the amount of data that can be stored and processed and also on the number of individual different values (resolution) a single datum can have.

8 bit = 256 individual values. 16 bit = 65535 individual values etc.

The Sensors that measure the amount of force you excert on a key are digital, the modeling algorithms and sound processing are digital so both have a fixed upper bound on the resolution or 'number of steps' that they can discern. The digital to analog converter (that converts the numbers into sound waves) also is digital.

Since your whole system deals with digital values and has technically 'lost' information due to the discretisation (conversion of a continuous signal of 'infinite' resolution into quantifiable digital values) of an analogue signal and the digital processing the digital to analog conversion cannot magically add the parts of the signal lost or never recorded/created.

Therefore a modeled piano also only has a finite number of individual sounds it can create. To overcome that one would have to build a truly analogue model of a grand piano - the grand piano itself.

Given that the number of individual samples is enough to match the number of individual steps the modeling algorithm can create then there would be no discernable difference between both approaches and you wouldn't be able to tell which is which.

Of course, the 128-step resolution of MIDI kicks in way before those other factors, I'm guessing.

Not nearly as limiting as you describe. For instance, audio quantized in both time and instantaneous level can be perfectly reconstructed in the bandwidth limited continuous domains with the only result of the quatization being some residual SNR. Done correctly it won't cause a perceptual loss of any kind.

Many oversampled converters modulate a single bit!

You lot have completely lost me there.....my V beckons, methinks.

That's just arguing semantics. We could quarrel about the exact definition of 'residual SNR' or '(no) perceptual loss of quality' all day, technically you lose information since you essentially cut off everything above a certain arbitrarily defined frequency threshold. That's why I used the term 'technically'.

We could argue if the loss of information is significant or noticeable but since we're debating the supposed superiority of modeling vs. sampling on the basis of better recreation of sound that would be a slippery slope.



Yeah, that's why they use oversampling the number of bits doesn't really have any significance here since the sampling technique is different.

But we're getting off topic here

Minus all the technical jargon, if a digital doesn't sound good (or, authentic) to your ears, then don't f*****g buy it!

Nigeth, you're wrong - there is no loss of information in modeled piano because "information" is created on the fly as needed not recreated as in sampled one. Same pointless discussion as with digital camera versus analogue...and analogue cameras are pretty much gone by now aren't they?

Yes, off topic. The perfect digital piano (if there were such a thing) need NOT exactly reproduce an acoustic one. If the two sound the same, then they are the same. Instrumentation might be able to distinguish them. But in the end only the ear and mind matter. If those are satisfied, the digital is as perfect as it needs to be.

So if the arbitrary cutoff frequency is above the threshold of hearing, instrumentation would know. But the ear would not. Similar arguments can be made for any parameter.

But the point is moot, for now. The ear CAN tell the difference between a digital and an acoustic. The question becomes: In what ways does the digital fall short? If someone knows the answer, then the question becomes: Will anyone do anything about those shortcomings?

Is quantization the problem? I doubt it.

Is the limited number of sample levels the problem? I think we're at (or near) the point where that ceases to be the problem.

In digital pianos, the short samples and looping of the native samples ARE a problem. But these are largely eliminated by the better samplers, and become irrelevant with the modelers.

So what's left? Does anyone know what the problems are? (I don't. Do the researchers know??)
And will something be done to make things better? (I can't. Can the developers??)

There are no answers... since digital technology has not been able to properly emulate acoustic phenomenon.*

*This includes every aspect of the way real acoustic piano strings interact with the soundboard, the case, harmonic resonance, etc.

Though there can also be instances where some people can hear a difference and others cannot.


I think the "acoustic space" may be the trickiest part. As I alluded to in another thread... you can record someone playing a piece on a spectacular acoustic piano with the best microphones, play it back through the finest speaker (or pair of speakers) you can get your hands on... and it still probably won't sound like there's a real acoustic piano in the room. And people expect digital pianos to sound indistinguishable from the real thing, through a couple of relatively inexpensive speakers yet. That's part of the cleverness of things like the Avant Grand, which use speakers to throw different aspects of the sound out of the cabinet in different ways. But you can't capture that effect on a recording, or in live performance through a PA. Maybe a binaural headphone system could come close.

But I think this means that, at least as far as a "slab" piano goes, playing it through regular speakers, there's really little hope of getting it to sound indistinguishable from a real piano... I think the more realistic goal is to try to get it to sound as if a real piano were being mic'd and played through speakers, which is not the same thing. But for a home console kind of application, a more sophisticated use of speakers could be more impressive than that.

I stopped short of saying that the speakers are the weak spot. But since you've raised that point ... I have to agree.

The list of problems must surely include other things. But the substitution of speakers for a soundboard acted upon by strings must be among the biggest of the shortcomings.

I'm sorry if we have confused you. So let me try to explain. Any sound, speech, music, noise is really a combination of sound waves.

A wave has two properties, amplitude i.e the power or volume of the wave and frequency i.e how 'fast' or how often the wave vibrates or changes from highest amplitude to lowest amplitude in a defined amount of time.

Both of those values, amplitude and frequency, can technically be arbitrarily large or small and the wave is a continuous thing (meaning that you won't find any pauses or gaps when you record it).

Both scales, amplitude and frequency, are of infinite resolution i.e you'd only be able to measure one or the other perfectly if you had a measuring instrument with infinite resolution (why exactly that is would lead us to far off the basic explanation).

A computer or digital system can only store a finite amount of information though, so it is only able to store a representation of your sound that 'kind of' resembles the original.

The computer can only assign a finite number of levels for the representation of the amplitude and it can only assign a finite number of levels for the representation of frequency.

That's true about all kinds of data a computer processes by the way.

Luckily there is a physical principle at work that a guy named Shannon discovered.

He discovered that one can 'perfectly' recreate a wave from its digital representation (both amplitude and frequency) if you measure it more than twice per period. If you do it more than twice per period then the exact time of your measurement doesn't even matter you can always recreate the waveform from the data.

So if I wanted to recreate a waveform with the frequency of 1 Hz (1 period per second) I'd have to measure the amplitude of the wave at least twice per second. If I wanted to measure a signal of 20,000 Hz or less I'd have to measure 40,000 times per second and so on.

That process is called discretization.

That's why CD's are mastered with a sampling frequency of 44,100 Hz for example so that you can store and recreate sounds up to about 20,000 Hz which is the limit most people can hear.

Turning that data back into sound is basically the same process in reverse.

You also have to measure the amplitude of the wave (loudness, volume, power) and you also only have a finite amount of space for that. Therefore the amplitude is also 'discretized'.

This process is 'lossy' though. If your amplitude is between two digital levels (and it can be since it has 'infinite' resolution in amplitude) you have to 'map' it to the nearest lower or higher digital level. So your measured amplitude is slightly lower or slightly higher than the original.

If you recreate sound from the digital value you'll get some additional noise that stems from that mismatch.

Consumer electronics often uses 44,1 kHz at 16 bit which means that every wave up to an upper limit of about 20 kHz is sampled and there are 16 bit or 65535 (2 to the 16th) different values available to store amplitude information.

There's an additional catch though. In order for that process to work at all you'll have to 'cut off' all of the frequencies above the maximum sample frequency. Otherwise you'll get a very unfortunate effect called aliasing.

Basically, since the converter only measures with say 20 Hz (20 times per second) he won't be able to discern a signal that has a frequency of 10 Hz for example from one that has twice/three times/four times/ the frequency.

It doesn't measure quickly enough and so all of the harmonics over the sampling threshold of 10 Hz (remember measure twice per period) look like waves below the sampling threshold. This can have some very unintended results if you play it back later.

To prevent this you'll have to prevent all of the frequencies above the sampling threshold from even entering the ad-converter.

If you have analog to digital to analog conversion (you record something digitally and then play it back later) this is achieved by using a low pass filter (a filter through which only waves up to a certain frequency can pass through) that is inserted before the analog to digital converter so that the AD-converter only 'sees' frequencies up to a certain value.

So now you have a signal that uses a finite number of levels to store the amplitude and which can only store signals up to an arbitrary frequency threshold.

Everything else is basically 'lost'.

Even if you don't work with an analog input signal but instead create the signal in a computer (for example with some sort of modeling) the basic principle stays the same. You only have a finite (e.g 16 bit) resolution for the amplitude and only a finite (say up to 20 kHz) resolution for the frequency information.

For it to be different would require a computer with infinite storage and the capability to handle an infinite amount of data in a finite amount of time.

Therefore there is a finite number of 'outputs' or steps you can create on a digital system.

There is a lively debate at which point the amount of useful information that is lost is small enough so that people won't notice it. Some say you'll always notice, it some say it's at 24 bit/96 kHz, some say even the invasive 'lossy' encoding of mp3 isn't really noticeable.

There are additional technical limitations in current systems for example midi only offers 7 bit (128 levels) in the standard implementation.

A modeling algorithm 'recreates' the sound of an acoustic within the limits of the digital system and the fidelity of the digital to analog conversion.

This recreation is an approximation that leaves out 'information' the real instrument would provide but the digital one cannot due to technical limits. So there certainly is a loss of information.

The only real argument is whether or not that approximation is close enough to the original so that a human would be unable to notice said difference.

20 years of the record vs. CD argument have told me that people will never agree on what constitutes "unnoticeable' or 'good enough' so good luck with that.

I'd argue that if the real criterion is 'you can't notice the difference' then you can't really argue that modeling is better than sampling or vice versa, At some point both technologies will cross into the realm of 'can't notice the difference' at which point the whole argument ceases to matter.

Which is my point because if the quality threshold is 'you won't notice the difference' then there can be no preference of modeling or sampling since both technologies have the potential to achieve said goal.



For modeling the answer is most probably that the algorithms are much simpler than they'd need to be in order to stay below a certain threshold with regards to computational power and price.

Modeling the complex interactions of all of the parts of an acoustic is a very hard and computationally expensive math problem. I'd guess that certain relationships and interactions between the different parts aren't even entirely understood yet and can't therefore be modeled at all.

So I'd guess there is still some research required and enough processing power to do all of the differential equation and fourier analysis work.

For sampling the biggest issue is storage space ( a single sample set for one acoustic grand can now use up to 12 gigabytes of disk space) and bandwidth to store an even greater number of samples for different velocity levels.

And one problem I can't really put a finger on but someone else in this thread posited also.

Somehow even the best sampled piano reproduced over the best sound equipment somehow doesn't feel like the 'real deal'. There is probably a sensory element that is not reproduced by the recording.

Nigeth,
You haven't quite explained why you think that modeled DP sound is no different from sampled DP sound, when the latter is derived from so few samples per note, whereas the former is generated from keystrike, with no preset number. Unless you're saying that the modeled sound also originates from no more than the same number of different possible sounds per note, only that it's computer-generated rather than from a recording, which I don't think is what you mean.

Any pianist worth his salt can produce more variations of dynamics - each with its own degree of overtones generated - than the number of actual sampled notes as recorded. And that's just for one note. When notes are combined, generating various resonances with different weighting for each individual note within say, an eight note chord (which any pianist worth his salt can do), a modeled DP runs rings over any sampled DP which just sounds like 8 notes all distinct from each other, rather than intermingling and producing different timbres depending on how the pianist voiced (in the classical sense) that chord, and what preceded it, like the decaying notes of another chord. In other words, the modeled sound produces a pretty convincing analog-like representation of an acoustic piano with all the blurring and clashing sounds and overtones; the sampled sound is sterile, and quite unlike what happens in the real thing. Even if, according to you (if I understand right), the modeled sounds have no more different sounds available than sampled sounds.

A CD obviously doesn't give the complete original sound recording, being a digital representation, just like a digital photo compared to a slide film, which I presume is your analogy above; but the information loss on CD (but not MP3) is to all intents and purposes undetectable by the human ear (though I'm sure most people can tell the difference between a digital photo and the smoother slide film photo, even if the former has lots of pixels), but the difference between modeled and sampled DP sound is easily detectable (even from just one struck note, even disregarding the looping that occurs with sampled DPs).

In that sense even real acoustic instruments are experiencing loss of information in real world due to weather, temperature, humidity and a distance from the listener....