Welcome to the Piano World Piano Forums
Over 2 million posts about pianos, digital pianos, and all types of keyboard instruments
Join the World's Largest Community of Piano Lovers (it's free)
It's Fun to Play the Piano ... Please Pass It On!

SEARCH
the Forums & Piano World

This custom search works much better than the built in one and allows searching older posts.
(ad 125) Sweetwater - Digital Keyboards & Other Gear
Digital Pianos at Sweetwater
(ad) Pearl River
Pearl River Pianos
(ad) Pianoteq
Latest Pianoteq add-on instrument: U4 upright piano
(ad) P B Guide
Acoustic & Digital Piano Guide
PianoSupplies.com (150)
Piano Accessories Music Related Gifts Piano Tuning Equipment Piano Moving Equipment
We now offer Gift Certificates in our online store!
(ad) Estonia Piano
Estonia Piano
Quick Links to Useful Stuff
Our Classified Ads
Find Piano Professionals-

*Piano Dealers - Piano Stores
*Piano Tuners
*Piano Teachers
*Piano Movers
*Piano Restorations
*Piano Manufacturers
*Organs

Quick Links:
*Advertise On Piano World
*Free Piano Newsletter
*Online Piano Recitals
*Piano Recitals Index
*Piano Accessories
* Buying a Piano
*Buying A Acoustic Piano
*Buying a Digital Piano
*Pianos for Sale
*Sell Your Piano
*How Old is My Piano?
*Piano Books
*Piano Art, Pictures, & Posters
*Directory/Site Map
*Contest
*Links
*Virtual Piano
*Music Word Search
*Piano Screen Saver
*Piano Videos
*Virtual Piano Chords
Page 7 of 8 < 1 2 3 4 5 6 7 8 >
Topic Options
#1885327 - 04/23/12 10:13 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
dmd Offline
1000 Post Club Member

Registered: 04/15/09
Posts: 1875
Loc: Pennsylvania
Well, I have been browsing through some of this conversation but will not pretend to be ready to jump in at the level you guys are working at ... but I would like to throw one thought into this.

If you find it to be trivial or not significant I encourage you to just ignore it completely ... lol ... but anyway here is what I think ...

If I am a pianist (of any skill level) and I strike a key and am expecting a particular sound to eminate from that piano ....

No matter how many sampling levels have been provided in that instrument ...

It is entirely possible that the sound I expect is not within the samples and I will be disappointed.

Conclusion: Sampling technology in digital pianos will never result in the same spectrum of tones as an acoustic piano.

On the other hand, I feel intuitively that modeling has no such limitations.
_________________________
Don

Current: ES7, Focusrite Scarlett 2i4 audio device, SennHeiser HD555 Phones, Focal CMS 40 Powered Monitors, Ravenscroft275, Ivory II American Concert D, Pianoteq 5

Top
(ads) Sweetwater / Roland
Special Financing on Digital Keyboards

Click Here


#1885368 - 04/23/12 11:45 PM Re: Sampled VS Modelled Pianos ??? [Re: sullivang]
Macy Offline
500 Post Club Member

Registered: 03/09/10
Posts: 611
Originally Posted By: sullivang
@Macy,
With respect, after having read your reply to Voxpops, I still don't think you have fully understood what I have been trying to say.

Adding the noise to the MIDI velocity is simply to restore the randomness that the quantisation has removed.

I keep thinking we understand each other, and then I guess we don't.

What randomness did the quantization remove? The quantization won't remove any significant amount of the players randomness if the quantization size is chosen so that the quantization noise is negligible compared to the player's random velocity noise. The quantization needs to be much smaller than the players randomness. I have said that over and over again.

So what randomness are you talking about that is removed by quantization?

Originally Posted By: sullivang
There are two issues here:
1. How many steps does it take to sample a human's intended velocity with sufficient accuracy?

Maybe this is where we are confusing each other. You can never know the player's intended velocity more accurately than the size of the player's velocity noise. If the players noise is +/- 5 current quantization steps than you can never know more accurately than +/-5 current quantization steps and only that if the new quantization size noise is zero (i.e. infinitely small new quantization step size). Any new finite size quantization will add additional inaccuracy to knowing the players intended velocity. So what accuracy do you want? What is sufficient to you?

Originally Posted By: sullivang
2. How many usable/interesting steps is the instrument able to reproduce?

Obviously you would have to have more usable sound generator steps than velocity quantization steps, else you just feed the velocity values to the generator without modification. Beyond that, who's to say if more are interesting or not?


Edited by Macy (04/23/12 11:58 PM)
_________________________
Macy

CVP-409GP, Vintage D, Ivory II GP's & American Concert D, True Keys American D, Ravenscroft 275, Garritan Authorized Steinway, Alicia's Keys, EWQL Pianos, MainStage, iPad/forScore/PageFlip Cicada, Custom Mac MIDI/Audio Software Design, Macs Everywhere

Top
#1885380 - 04/24/12 12:12 AM Re: Sampled VS Modelled Pianos ??? [Re: Macy]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
Originally Posted By: Macy

What randomness did the quantization remove?


Some of the randomness of the player's natural variability, because the quantisation produces a velocity signal that can only take on discrete values. The quantisation removes intermediate steps from the player's performance, which may be important for any given instrument. I am saying that with experiment, we can probably arrive at some sensible minimum resolution that is required to represent a human's velocity signal. This will allow for optimum use of MIDI bandwith, storage of MIDI files, etc. Any more resolution, and we start to capture the signal with more resolution than makes sense. If this signal happens to be fed to an instrument that really does produce noticable tonal shades between steps available in our MIDI velocity, as I've said before, it may well be that simply adding a random peturbation to the signal (amplitude +/- 0.5 steps) will produce a result that feels and sounds the same to the player, as if we had quantised with higher resolution to begin with. We've already sampled with a very high resolution (as far as the human is concerned), yet, the player may notice that the intermediate steps are never ever produced due to this quantisation. Just because a player can't always hit a certain velocity on demand doesn't mean that we can omit it altogether!

Quote:
The quantization won't remove any significant amount of the players randomness if the quantization size is chosen so that the quantization noise is negligible compared to the player's random velocity noise. The quantization needs to be much smaller than the players randomness. I have said that over and over again.


Agreed, but it may well remove enough randomness to be a problem for any given instrument. Take another extreme example - a piano that responds to 10,000 discrete velocities, yet we are feeding it with 127. The player may not notice any problems, except for one very important quirk with this piano - it makes the sound of a chirping bird at velocity 9567, and ONLY that velocity. The player has learnt from experience when playing on a high-res controller, that if he plays at around a certain strength, he can sometimes, but not always, invoke the chirping bird sound. If they now switch to the low-res (127 step) controller, and none of the available 127 steps correspond to the piano's internal velocity of 9567, the bird sound will never be produced, and this is a serious problem.

Quote:
You can never know the player's intended velocity more accurately than the size of the player's velocity noise. If the players noise is +/- 5 quantization steps than you can never know more accurately than +/-5 quantization steps and only that if the quantization noise is zero (i.e. infinitely small step size). Any finite size quantization will add additional inaccuracy to knowing the players intended velocity. So what accuracy do you want? What is sufficient to you?


Let me come back to my very extreme example:

Again, let's say we are designing a MIDI keyboard for a particular player. Upon very careful analysis, using a very high resolution MIDI keyboard, we determine that no matter what velocity this person attempts to play, the resulting velocity is a random number between and 1 and X, X being a very high number (the number of steps of our test bed keyboard). Having performed this prior analysis, we decide that a MIDI keyboard that will be suitable for this player can consist of keys that are simple on/off switches. When the key is pressed, it triggers a random number generator. There is no need to sample this players analog velocity signal with any more than 1 bit (on/off switch), because we know that we can synthesize his random noise to produce an impression that is indistinguishable from his actual noise. The resolution with which we generate this random number will depend on what we want to do with it. If we want to play a piano, and be able to (randomly!) stimulate all possible tone gradations, we may need, say, 1000 different steps. If we want to stimulate a harpsichord, we can reduce the resolution, to, say, 10 steps, because we've determined that the 10 steps are so similar not to warrant any intermediate steps.

Do you agree with this so far? In this case, the answer to your question is that only one bit is required to capture the player's velocity signal with sufficient accuracy, because of our prior analysis of this player. I want enough accuracy to represent what the player is capable of doing determistically.

Anyway, if you can at least agree with this simple example, maybe we can move on, by increasing the precision of the player. It will not be as neat and tidy as this example, but I still think my overall idea has merit.

Quite a few edits in the above - sorry.

Greg.


Edited by sullivang (04/24/12 05:40 AM)

Top
#1885466 - 04/24/12 05:31 AM Re: Sampled VS Modelled Pianos ??? [Re: sullivang]
Macy Offline
500 Post Club Member

Registered: 03/09/10
Posts: 611
Originally Posted By: sullivang
Originally Posted By: Macy

What randomness did the quantization remove?


The randomness of the player's natural variability, the extent of which is determined by prior experiment with humans, using statistical analysis. I am saying that with experiment, we can probably arrive at some sensible minimum resolution that is required to represent a human's velocity signal - this signal excludes the natural random variability. This will allow for optimum use of MIDI bandwith, storage of MIDI files, etc. Any more resolution, and we start to capture random human noise, which is not important, because we can easily synthesize it if we want to.

You are trying to violate information theory. You can't remove the player's natural variability (what I call player noise) from your signal by quantization. You can't create a signal that represents only what he intended and excludes his natural random variability. No matter what size quantization you choose, the player variability (what I call player noise) will always trigger the output of more than one quantization level for some intended velocities. That's fundamental. Every quantizer has a threshold level that triggers it to output its next value. So whenever your intended velocity is within an interval the size of the peak to peak noise around the threshold level it will randomly trigger the quantizer to output a different value. As the quantization size is increased, multiple output values are produced less often, but the error between the intended velocity and the larger quantization interval values is bigger, and hence the players variability (the player's noise) is preserved and can not be removed.


Originally Posted By: sullivang
Again, let's say we are designing a MIDI keyboard for a particular player. Upon very careful analysis, using a very high resolution MIDI keyboard, we determine that no matter what velocity this person attempts to play, the resulting velocity is a random number between and 1 and X, X being a very high number (the number of steps of our test bed keyboard). Having performed this prior analysis, we decide that a MIDI keyboard that will be suitable for this player can consist of keys that are simple on/off switches. When the key is pressed, it triggers a random number generator. There is no need to sample this players analog velocity signal with any more than 1 bit (on/off switch), because we know that we can synthesize his random noise to produce an impression that is indistinguishable from his actual noise.


What an interesting example. My first reaction was to simply dismiss it on the grounds that it didn't apply to the piano problem because once the piano keys were replaced with simple switches the player's input to the system was strictly digital, not analog velocity like the real piano situation. And everything we have discussed here centers on how to quantize a players analog input to the system. How much noise there is in the player's analog velocity input, how big should the quantization intervals be for his analog velocity input compared to the noise from the variability of his analog velocity input, etc. So if the players input to the system is strictly a digital on-off input, it doesn't apply to our problem at all. There has to be quantization of a player's analog input, for all these other factors to come into play.

But ...... upon further reflection, I realized the player's input to your hypothetical system is still analog as long as he is pushing the switches to trip them, and thus this interesting example can under that circumstance apply to our situation. So as long as there is an analog input to the switches, and even as simplistic as this concept of enabling a random number generator may be compared to playing a piano, we can still learn a great deal from it.

In this example the players input to the simplified keyboard with just on-off switches is still an analog velocity (force) to trip those switches. Therefore, the switches must have some quantizing threshold velocity (force) value that has to be overcome before they turn on. In other words, the keyboard is still quantizing the players analog input but it only has two output levels, off and on.

So let's go back and see if the original system you described, where the player always creates a random velocity output no matter what velocity he attempts to play, creates the same output in the 1-bit keyboard quantization example. In order for them to be the same, "indistinguishable from his actual noise" in your words, the 1-bit keyboard must always produce a random velocity value every time the player presses a key (switch). But of course, it doesn't. Because if the random velocity of the player is less than the quantization threshold of the 1-bit switch, the random number generator is not enabled at all. It produces nothing, no random numbers at all until the quantization threshold is exceeded.

So the output of the 1-bit keyboard device is very distinguishable, rather than indistinguishable, from the output of the initial piano keyboard device. They are not the same. The initial keyboard with its very high resolution keyboard always produced a random number output, while the 1-bit low resolution keyboard randomly produces nothing at all. So the concept of quantizing an analog input and then adding random noise to the digital output to duplicate the random noise of the analog input signal didn't work because of the analog quantization, as I have been saying. The result is extreme in this case - randomly no output at all, because of your hypothetical example of 1-bit (on-off) input quantization. Its an extreme example of the general case of quantization error (noise) modulating the output, and it's the players input noise (which can not be eliminated by quantization as I discussed above), that creates the quantization error and is hence preserved. So your example illustrates my point at the beginning of this post.

With this, I must bow out of this discussion. We are starting to just go around in circles. It would be fun to continue this conversation sitting in a bar having a few beers. Either there would be a sudden "I get it" moment or eventually we probably wouldn't care any more. But it is consuming too much of my time (and probably yours?), and I'm sure doesn't interest anyone but you and me anymore, to keep rehashing this here. So respectfully, and for no other reasons, I have to let this go now.
_________________________
Macy

CVP-409GP, Vintage D, Ivory II GP's & American Concert D, True Keys American D, Ravenscroft 275, Garritan Authorized Steinway, Alicia's Keys, EWQL Pianos, MainStage, iPad/forScore/PageFlip Cicada, Custom Mac MIDI/Audio Software Design, Macs Everywhere

Top
#1885489 - 04/24/12 06:28 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@Macy:
In my one-bit keyboard, the switch only has two states - on, and off. The minimum force required to switch it ON is exactly the same as the force required to produce the minimum velocity of the high resolution keyboard. When it switches ON, it triggers the random number generator.

Greg.


Edited by sullivang (04/24/12 06:31 AM)

Top
#1885493 - 04/24/12 06:45 AM Re: Sampled VS Modelled Pianos ??? [Re: sullivang]
Macy Offline
500 Post Club Member

Registered: 03/09/10
Posts: 611
Originally Posted By: sullivang
@Macy:
In my one-bit keyboard, the switch only has two states - on, and off. The minimum force required to switch it ON is exactly the same as the force required to produce the minimum velocity of the high resolution keyboard. When it switches ON, it triggers the random number generator.

Greg.

Greg, I can't keep responding. I got to move on. I understood that in your example, and as I said, that makes the example not apply to the piano problem because there would be no quantization of an analog input signal. So I simply made the example apply by making the switch work more like the piano problem. It does take force to trigger a switch. Therefore, it does have a quantization threshold. And you said the original keyboard had very high resolution, so to make the example have generality, I simply hypothesized a switching threshold somewhere within the velocity range of the player, rather than at the minimum range, which has no generality.

Now it 3:45 am here. I earlier wrote you a PM. Best wishes and good night.
_________________________
Macy

CVP-409GP, Vintage D, Ivory II GP's & American Concert D, True Keys American D, Ravenscroft 275, Garritan Authorized Steinway, Alicia's Keys, EWQL Pianos, MainStage, iPad/forScore/PageFlip Cicada, Custom Mac MIDI/Audio Software Design, Macs Everywhere

Top
#1885528 - 04/24/12 07:49 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@Macy,
I'm not forcing you to reply, and if you don't, others can step in. I'm learning a lot from you, and I appreciate it.

I agree that my example is very contrived. The player is always able to overcome the switching threshold, yet, once he does, he then produces perhaps a much wider range of forces, at random. Not very realistic, however, it's still a possible behaviour for this strange player. Due to our prior analysis of this player, we have determined that no quantization is required. So, we do not in fact have to sample the noise with any precision at all during a performance - we can synthesize it. If we had not done the prior analysis, then we could simply sample this player with the finest resolution that any given instrument is capable of.

The reason I've been going on about this so much is that I feel that two different issues have been jumbled up in people's minds, and it has led to very vague discussions. The resolution that a player is capable of performing at may be very different to the resolution of the instrument. The main point I'm trying to make is that if we have a MIDI signal that is indeed able to capture a person's velocity with enough precision to represent his ability, this precision (number of steps) may still be insufficient to trigger all available tonal colours in the instrument, because the quantisation has omitted the intermediate steps that the instrument is capable of. (please see my example two posts ago re: the chirping bird piano). I think this is what others have been saying all along. If a player can produce a given tonal colour sometimes, this tonal colour needs to be retained, if the player likes being able to produce it sometimes.

Adding +/- half a step of noise to the output (at the resolution of the instrument, or higher) seems reasonable to me, to ensure that the intermediate steps are invoked at random. The input voltage could have been any of these new output values, so we're unlikely to be doing any serious damage to the signal. For example, if the input voltage can vary in a smooth manner between 1V and 10V, and we quantise to 10 steps, the possible output values will represent 1.5V, 2.5V, ....8.5V, and 9.5V, with the quantization error being +/- 0.5V. So, for step 1 (1.5V), we're really saying that the input signal was anywhere in the range 1V to 2V, and so on. Adding a +/- half a step (+/- 0.5V)random noise signal to the output will hopefully restore the randomness that the quantisation removed.

Unless we're willing to run any tests on humans to determine the required precision to capture there abilities independently of any musical instrument, then we have to assess them with each instrument seperately, because we don't know how precisely a player can play.

Greg.

Top
#1885583 - 04/24/12 10:00 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
ando Offline
3000 Post Club Member

Registered: 11/23/10
Posts: 3588
Loc: Melbourne, Australia
Greg, I am not in agreement with your theory for reasons I think I made clear in earlier posts, but as a matter of practicality: If you were going to write randomness in between the current quantised steps, they would fall on predetermined quantised steps of their own, right? You are effectively creating a higher output resolution. For a sampled system, how would you want to implement this? Would you be advocating more samples? More volume levels? More filtering levels? I'd be surprised you would think it's worth doing all that for the sake of some randomness when you could achieve the same thing without randomness and actually use higher resolution MIDI on the input side too - and actually have it accurately coupled to the player. The only reason I can think of for not doing it this way is if you wanted to retain the old MIDI standard for the sake of cost. On a technical level, hi-res beats random "dithering" - or whatever we are calling it now.

Top
#1885599 - 04/24/12 10:21 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@ando:
Let's say we are indeed limited to 127 steps for our MIDI signal. The way I would do it is to test the instrument very carefully, to determine the finest gradation of tonal colour that a) is audible, and b) of interest. Let's say that I discover that there are in fact 511 different tonal shades of interest. (in reality, I may discover that the total number of useful shades is much less than 511, however the smallest step size of interest is equivalent to the step size that results in dividing the velocity signal into 511 equal sized steps). 511 steps can be represented by a 9-bit quantisation.

To allow the player to experience these additional tone shades, we can simply divide each of the 127 steps in to four sub-steps, increasing the "fake" resolution of the output signal. (as you said). The addtional sub-steps are created by using a random number generator that produces 4 values, and we use this to select one of four sub-steps. We're not degrading the original signal, because all these four new values lie within the range of the original step size.

I totally agree - the better way is to simply increase the resolution of the MIDI velocity, and do away with this noise superposition idea. However, if we do this, this does not mean I am admitting that the player can necessarily produce these new intermediate steps in a deterministic fashion, and the overall experience for the player may be very hard to distinguish from using the 127 steps + noise idea.

As to how to create the sound, we simply use the new, higher resolution (511 step) signal to control the soundsource. If it's a Disklavier, the digital signal will somehow control the speed of the hammer. If it's a synth, the velocity will somehow control aspects of the sound - it would depend on the preset and how the synth is designed. All that really matters is that the sound generator somehow generates 511 different tonal shades.

Greg.


Edited by sullivang (04/24/12 10:30 AM)

Top
#1885645 - 04/24/12 12:07 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
Correction to my previous post.

I can now see what Macy was trying to tell me earlier. (sorry Macy!) Adding the random noise does increase the quantization noise, despite the fact that the new output values all lie in the same range as the original input signal. This is an unfortunate tradeoff. It appears that in our example (127 + 4 steps of noise), the original quantization error is +/- half a step. Adding the noise, the new error is +/- 7/8ths of a step.

For very well behaved instruments (such as a fine concert grand), I think that adding this noise could be detrimental, and we simply have to increase the MIDI resolution proper in order to make any worthwile improvement. At the very least, we have to take more care than I thought. If we think we can safely sacrifice the resolution, then no problem.

For instruments that exhibit some more quirkiness, the ability to at least invoke the missing tonal shades may be worth the tradeoff of the loss of velocity resolution.

Greg.


Edited by sullivang (04/24/12 05:02 PM)

Top
#1886007 - 04/25/12 12:06 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
I've been absolutely racking my brain about how to determine what resolution we need to use to capture the velocity signal, and I think I can have a crack at defining the problem mathematically.

If a human tries to play a given velocity repeatedly, the resulting output velocities will probably follow some kind of bell curve distribution, centered on the desired velocity. We need to reproduce this curve, however the accuracy with which we need to reproduce it will depend on how strong the feedback is to the player. I.e - it depends on how distinguishable each velocity level actually is.

So, going back to my 1-bit velocity sensor - the reason we only needed one bit is because the player plays with an infinitely wide bell curve. (i.e in a purely random fashion). All we have to do is detect the key press, and then generate a pure random number.

Now, let's say the player is a bit more accurate. The bell curve is still very wide, but it does taper a bit. The centre of the curve is on the target velocity, but the curve gently tapers down each side. To reproduce this curve with sufficient accuracy, we run another experiment, and determine that we can quantise the velocity signal to just four different values, and add white noise peturbation to the four steps, so that all output velocities are still reproduced. The actual output probability distribution curve for any target velocity will be a very coarse step-ladder, but because the curve we are trying to reproduce is so flat, this is ok - it's accurate enough. The reason it's a step ladder is because the white noise produces the same number of occurrences, on average, for all velocities in a given quantisation step. We have to use white noise, because we have no idea what part of the bell curve we're on, for any given input velocity. All we can do is choose the appropriate quantisation step size such that the quantisation + white noise reproduces the input bell curve distribution closely enough such that the player can't tell the difference.

As the player becomes more accurate, the bell curve becomes narrower and narrower, and of course we need to quantise with finer gradations to reproduce this curve.

I may need to do a diagram. wink

Greg.


Edited by sullivang (04/25/12 12:15 AM)

Top
#1886405 - 04/25/12 03:37 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
My so-called "1-bit" velocity sensor is, in fact, a zero bit velocity sensor.

A 1-bit sensor would have two possible outputs. (binary number 0, and 1, for example) It would be sensing the two halves of the velocity range. The system would be able generate two MIDI velocities, with the values 1, and 2. (there is no MIDI velocity 0 - that is reserved for a Note-Off).

For a player that plays purely randomly, we don't need two values - we only need one value. We'll get no MIDI event if the player doesn't play a note (or if they play too softly to trigger the note). Whenever they play a note, we'll get a MIDI Note-On event with a velocity of 1. (or any velocity value we feel like, because there is no longer any velocity information at all!)_

If you like, we could formally design our 0-bit sensor, to act like a proper analog to digital converter. It would be sensing one velocity range with a range of 100%. All velocities would be mapped to the centre value (50%, with an error margin of +/- half a step, so the error margin would be +/- 50%!). I.e - the output of this sensor would always be 50% +/- 50%, in similar way to how an analog to digital converter maps the voltages in each range to the centre value. In this case, there is of course no useful information because we already know that if a note is played, the velocity will be in the range we can sense.
In our case of the player that generates purely random velocities, we take this 50% value and generate white noise around it, with an amplitude of +/- 50%.

With only one range to sense, though, there's no need to go to all that trouble. We only need the switch, which is also required for the high resolution normal sensor!). We only need to know when a note is played.

Greg.


Edited by sullivang (04/25/12 04:32 PM)

Top
#1886451 - 04/25/12 04:58 PM Re: Sampled VS Modelled Pianos ??? [Re: dmd]
R_B Offline
500 Post Club Member

Registered: 09/03/09
Posts: 503
Originally Posted By: dmd
Well, I have been browsing through some of this conversation but will not pretend to be ready to jump in at the level you guys are working at ... but I would like to throw one thought into this.

If you find it to be trivial or not significant I encourage you to just ignore it completely ... lol ... but anyway here is what I think ...

If I am a pianist (of any skill level) and I strike a key and am expecting a particular sound to eminate from that piano ....

No matter how many sampling levels have been provided in that instrument ...

It is entirely possible that the sound I expect is not within the samples and I will be disappointed.

Conclusion: Sampling technology in digital pianos will never result in the same spectrum of tones as an acoustic piano.

On the other hand, I feel intuitively that modeling has no such limitations.



Depends how WIDE the achieved spectrum is - and/or how granular your perceptions are within it.

I can imagine very much faster mass storage devices that are also very much larger, "smart" caching and look-ahead. Smart caching includes directory caching.

"Enough" AI in a sample player COULD enable it it anticipate "musically" and have just what you want just when you want it - or even slightly before you want it, just in case you "go there", with several alternate "there" places.

For the near and mid term I thing some hybrid solution offers the best compromise.
Even a "perfect" physical piano is a large wooden box full of compromises, we live with them (-:

Top
#1886477 - 04/25/12 05:41 PM Re: Sampled VS Modelled Pianos ??? [Re: Macy]
DazedAndConfused Offline
Full Member

Registered: 10/26/10
Posts: 198
Loc: Greenwich, London, United King...
Originally Posted By: Macy
Originally Posted By: Voxpops
Let's take your reasoning and apply it in the opposite direction. That would imply that the fewer the steps available the more likely I am to be able to play accurately, because I am much more likely to be able to strike at a particular velocity and therefore meet my feedback/restrike criteria. I think 10 steps should do it. wink


True the fewer the steps the more likely that you could hit the steps you intend out of the available steps. The problem is that too few steps will sound awful. Try 3 steps for instance. You won't have much expression because you have so few loudness/timbre choices to play for each note. So we increase the steps to give you more choices and your playing can sound better. Do we continue to 10,000 steps just because we could? At some point, the steps get so small we can no longer control our motor skills well enough to reliably play the steps we would like to play, even assuming we can somehow (consciously or sub-consciously) process the feedback from our playing with enough resolution to intend to play in those smaller steps. We simply can't do it.


But given the feedback we can try ... and always fail to hit it exactly ... and respond to that failure with the next note etc and out of that process of continuous adjustment comes expression and musicality. With insufficient feedback the process stalls. The ability to hit the exact step is unimportant. The motor control to get close enough to what you intended to play and the precise audible feedback that allows a player to respond is everything. 127 steps isn't nearly enough IMO.

Humans might not be very good at describing small differences between sounds but we are much better at detecting when two sounds are absolutely identical which is why I believe that in your cited Disklavier vs Sampled piano test it would be very easy to identify the sampled piano, especially if the first note was still decaying as the second one is triggered.

One would sound like a repeated recording. The other would sound like a piano.
_________________________
Currently working on:
Poulenc, Mouvements Perpetuels
Shostakovich, Prelude & Fugue no. 5
Beethoven, Sonata in F Op. 10 No. 2

Top
#1886560 - 04/25/12 08:00 PM Re: Sampled VS Modelled Pianos ??? [Re: DazedAndConfused]
R_B Offline
500 Post Club Member

Registered: 09/03/09
Posts: 503
Betamax vs VHS

Any takers ?

Top
#1886591 - 04/25/12 08:58 PM Re: Sampled VS Modelled Pianos ??? [Re: DazedAndConfused]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
Originally Posted By: DazedAndConfused

Humans might not be very good at describing small differences between sounds but we are much better at detecting when two sounds are absolutely identical which is why I believe that in your cited Disklavier vs Sampled piano test it would be very easy to identify the sampled piano, especially if the first note was still decaying as the second one is triggered.


The right way to test is to fit a standard 127 step keyboard to the Disklavier, and see if you can tell any difference. wink If you can, the there's no question that 127 steps is insufficient for you - I totally agree. Of course, the keyboard must be identical in every other way to the high resolution keyboard. (so we'd probably just use the high resolution keyboard, and quantise the high resolution MIDI down to 127 steps).

[EDIT:] Just by the way, I've looked at the "Dynamic Range" setting on some Pianoteq presets. The maximum I've found so far is 54dB. If we divide 54dB by 127, we get about 0.5dB. I.e - assuming every velocity step changes by the same amount, each step will be 0.5dB. The minimum change in sound level that can be readily detected by humans is considered to be 1dB. So, in terms of sound level alone (which is only part of the picture), 127 steps seems enough for a piano. (or at least Pianoteq, if a real piano has a higher dynamic range than this for any single note)

Greg.


Edited by sullivang (04/25/12 09:24 PM)

Top
#1886621 - 04/25/12 10:18 PM Re: Sampled VS Modelled Pianos ??? [Re: sullivang]
voxpops Offline
3000 Post Club Member

Registered: 03/20/07
Posts: 3050
Loc: Oregon
Originally Posted By: sullivang
The minimum change in sound level that can be readily detected by humans is considered to be 1dB.

Not necessarily:
"The minimum change in SPL required to give a detectable change in the loudness sensation (JND in sound level) is roughly constant and of the order of 0.2 - 0.4 dB in the musically relevant range of pitch and loudness."
From The Psychophysics of Loudness by David Worrall of the Australian National University
http://www.avatar.com.au/courses/PPofM/loud/Loud1.html
_________________________
Occasional author and inveterate ivory tickler:
http://www.amazon.com/author/richardspanswick

Top
#1886631 - 04/25/12 10:34 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@voxpops:
Very interesting. I can't remember where I got the 1dB figure from, but here's a page with the results of a few studies:
Human Hearing: Amplitude Sensitivity Part 1

and the author of this page says:
"I tend to use .75 dB to 1 dB when considering minimums."
(some of the referenced studies do report a smaller result, though)

Greg.

Top
#1886656 - 04/25/12 11:02 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
voxpops Offline
3000 Post Club Member

Registered: 03/20/07
Posts: 3050
Loc: Oregon
Greg, I'm no technical expert, but if Mr. Worrall is correct, then those most sensitive to sound changes (which would presumably include pro musicians) would require a minimum of 270 steps to cover a 54dB spectrum and not be able to detect the changes.

Obviously not everyone will be able to detect 0.2dB changes, but that would seem to be a not unreasonable standard to aim for when emulating the dynamics of a piano more accurately.
_________________________
Occasional author and inveterate ivory tickler:
http://www.amazon.com/author/richardspanswick

Top
#1886660 - 04/25/12 11:12 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@voxpops:
Agreed. I merely wanted to find a reference to support my 1dB figure.

What I am most certainly not saying, though, is that we would necessarily need to sample the player's velocity with that resolution, because I'm still of the opinion that we could do a very good job of synthesizing a human's natural variability.

Greg.


Edited by sullivang (04/25/12 11:15 PM)

Top
#1886670 - 04/25/12 11:32 PM Re: Sampled VS Modelled Pianos ??? [Re: sullivang]
ando Offline
3000 Post Club Member

Registered: 11/23/10
Posts: 3588
Loc: Melbourne, Australia
Originally Posted By: sullivang
@voxpops:
Agreed. I merely wanted to find a reference to support my 1dB figure.

What I am most certainly not saying, though, is that we would necessarily need to sample the player's velocity with that resolution, because I'm still of the opinion that we could do a very good job of synthesizing a human's natural variability.

Greg.


I don't like this approach. Let's say you play a note and it comes out on the loud side due to the randomisation in your system. You are aware of the loudness, so the next note, you try to back off a bit, but it goes the other way - that is, the randomiser shunts it lower, so then it's too soft. This example functions as a dynamic expander. The next time it might act as a compressor. You would have no confidence in the result of your next note(s). Your system is biased too much toward a listener who is decoupled from the player. Sure it might make it more interesting for a listener, but I don't believe you can assist the player by adding randomness to their input. All it can do is make no effect, or decouple them from their instrument.

Top
#1886671 - 04/25/12 11:37 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
anotherscott Offline
3000 Post Club Member

Registered: 02/20/10
Posts: 3226
Regardless of what the smallest perceptible volume difference is in decibels, remember that, with changing velocity, the piano sound is not merely changing in volume (if that were the case, we'd only have to sample one velocity), it is changing in other ways (timbre, envelope). The fact that it is changing in other ways may make it easier to hear smaller velocity differences than you'd be able to detect if the only difference were volume.

Top
#1886672 - 04/25/12 11:45 PM Re: Sampled VS Modelled Pianos ??? [Re: anotherscott]
voxpops Offline
3000 Post Club Member

Registered: 03/20/07
Posts: 3050
Loc: Oregon
Originally Posted By: anotherscott
Regardless of what the smallest perceptible volume difference is in decibels, remember that, with changing velocity, the piano sound is not merely changing in volume (if that were the case, we'd only have to sample one velocity), it is changing in other ways (timbre, envelope). The fact that it is changing in other ways may make it easier to hear smaller velocity differences than you'd be able to detect if the only difference were volume.

Good point.
_________________________
Occasional author and inveterate ivory tickler:
http://www.amazon.com/author/richardspanswick

Top
#1886681 - 04/26/12 12:00 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@anotherscott:
Agreed, and that's precisely why I said that the volume was only part of the picture in my original post! smile

@ando:
I agree with everything you say, however, I think I would able to synthesize it to sufficient accuracy that you simply would be none the wiser. If you can detect the difference with my system, I have failed.

Greg.


Edited by sullivang (04/26/12 12:00 AM)

Top
#1887132 - 04/26/12 07:47 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
My DAW (Ableton Live) has a velocity randomizer effect - it can generate noise centred on each velocity value, and the amplitude of the noise can be adjusted. (i.e +/- 1 step, 3 steps, 45 steps... etc). We could experiment with that, to see whether we notice the difference with different amplitudes of noise, with different instruments. (I can't do any testing at the moment) Haven't found a velocity "quantizer" yet, although that could easily be created with anything that allows per-step velocity mapping.

Suggested test: without the randomizer, train yourself for as long as you like, to try and strike any given velocity repteadly. Use the strongest form of feedback - look at the MIDI velocities on the screen. Then. turn on the randomizer, with a deviation of +/- 1, and see if you can notice any difference in the frequency with with you manage to hit that same velocity repeatedly. If you can notice a difference, then this means that you can't tolerate any less than 127 steps (for this specific test only), and in fact, you may need more than 127 steps. This is an extreme test - if you can't notice the difference this way, you will definitely not notice the difference using a piano sound, IMHO.

Thinking about it, another way to do the test is to do a statistical analysis of the results without the randomizer. If you strike each of the neighbouring velocities (+1 and -1) approximately the same number of times as the target velocity, we could then conclude that 127 steps is aready enough to represent your physical accuracy.

Greg.


Edited by sullivang (04/26/12 08:12 PM)

Top
#1887178 - 04/26/12 10:16 PM Re: Sampled VS Modelled Pianos ??? [Re: sullivang]
R_B Offline
500 Post Club Member

Registered: 09/03/09
Posts: 503
I read the whole thread yesterday, today I reflected on SOME of it.
The poor performer without a clue who can only make a couple of different velocities kept coming to mind (too close to home, perhaps I only make 7).

Then the path from finger strike to hammer strike, the piano's action.
I don't see any of that as "frictionless" and with friction comes stiction, so I wonder how many different velocities a hammer can actually achieve in a physical piano - even if the performer could achieve some arbitrarily HUGE number of different finger strike velocities, whether by skill or by accident.
A really STICKY action might result in some very small number, less "sticky" (on a "well played in" piano) could result in significantly more.
I have NO IDEA where this would scale relative to the MIDI standard of 127, or to 255, 1023, or 2047 (or 7).

This COULD be fairly easily measured, though I don't have the facility for it.

I guess it would be worth knowing though ?

Top
#1887875 - 04/28/12 06:14 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
I had a go at repeatedly striking velocity #64 on an old Yamaha PSR-230 keyboard. (cheap unweighted action). Note that the dynamic range was good - 12 to 127, and I had to play very strongly to reach 127. I was looking at a MIDI monitor throughout the test.

Here is the histogram:
http://i47.tinypic.com/2exufk9.png

127 steps might be more than enough for this particular keyboard and/or myself, in the specific (and very constricted) conditions of this test. Will repeat on a DP keyboard at some stage.

Greg.


Edited by sullivang (04/28/12 06:16 AM)

Top
#1887899 - 04/28/12 08:15 AM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
anotherscott Offline
3000 Post Club Member

Registered: 02/20/10
Posts: 3226
I think that difficult velocity control is specifically one of the shortcomings of many unweighted actions, and is one of the reasons they tend to be so poor for piano playing.

Top
#1888112 - 04/28/12 07:14 PM Re: Sampled VS Modelled Pianos ??? [Re: musicmad]
sullivang Offline
2000 Post Club Member

Registered: 07/05/09
Posts: 2207
Loc: Sydney, Australia
@anotherscott: I'm sure you're right, and as I said, I'll repeat the test on a DP when I can. In the meantime if anyone else would like to do the test and upload the MIDI file, I'll analyze it and upload the histogram.

Note that I wasn't 100% focused for the entire test. It was extremely boring.

Greg.

Top
#1893943 - 05/09/12 04:24 AM Re: Sampled VS Modelled Pianos ??? [Re: piano_shark]
galaksa Offline
Junior Member

Registered: 10/04/10
Posts: 9
Loc: Saint-Petersburg, Russia
Originally Posted By: piano_shark
Originally Posted By: musicmad

So what is the Grand Piano D4 ? If they got the sound from a Steinway D ?


And back to topic - sampled pianos are gonna die sooner or later,
modeled is the way to go. IMO V-piano and pianoteq are best piano simulation out there. (comparing to real grand of course)


V-piano and pianoteq only? There is another physical modeling technology that is worthy of attention. Its DRAKE by Generalmusic. As known Generalmusic is gone as a manufacturer, but their products are still available (Promega, RP & PRP pianos, RP-X module). My first encounter with the DPs was GEM RP-810 and I could not tear myself away from that natural sound. Much later I discovered that the piano was based on physical modeling. Now I have GEM RP-X piano module, it uses only 64MB of samples (50 voices sound list), physical modeling does the rest. I've heard V-piano performance and IMHO my GEM RP-X sounds no worse than a V-piano.
Here is a Fazioli Grand sound from my GEM RP-X:

https://www.box.com/s/41291d01f02d604bdfed

As for the samplers-romplers, all the worthy ones are for laptops only. But laptops tend to become obsolete. While the old familiar hardware friend will please you forever...

Top
Page 7 of 8 < 1 2 3 4 5 6 7 8 >

Moderator:  Piano World 
What's Hot!!
HOW TO POST PICTURES on the Piano Forums
-------------------
Sharing is Caring!
About the Buttons
-------------------
Forums Rules & Help
-------------------
ADVERTISE
on Piano World

The world's most popular piano web site.
-------------------
PIANO BOOKS
Interesting books about the piano, pianists, piano history, biographies, memoirs and more!
(ad) HAILUN Pianos
Hailun Pianos - Click for More
ad (Casio)
Celviano by Casio Rebate
Ad (Seiler/Knabe)
Knabe Pianos
Sheet Music
(PW is an affiliate)
Sheet Music Plus Featured Sale
(125ad) Dampp Chaser
Dampp Chaser Piano Life Saver
(ad) Lindeblad Piano
Lindeblad Piano Restoration
New Topics - Multiple Forums
Modern Jazz Pianists and Compositions
by b4ggy
12 minutes 4 seconds ago
Interview with Leon Fleisher
by vlhorowitz
Today at 04:27 AM
Interview with Leon Fleisher
by vlhorowitz
Today at 04:26 AM
Interview with Leon Fleisher
by vlhorowitz
Today at 04:26 AM
First sonata
by amt976
Today at 03:38 AM
Who's Online
85 registered (Alexancho, amt976, Alux, Al LaPorte, 36251, 21 invisible), 1146 Guests and 19 Spiders online.
Key: Admin, Global Mod, Mod
Forum Stats
76384 Members
42 Forums
157913 Topics
2319124 Posts

Max Online: 15252 @ 03/21/10 11:39 PM
(ads by Google)

Visit our online store for gifts for music lovers

 
Help keep the forums up and running with a donation, any amount is appreciated!
Or by becoming a Subscribing member! Thank-you.
Donate   Subscribe
 
Our Piano Related Classified Ads
|
Dealers | Tuners | Lessons | Movers | Restorations | Pianos For Sale | Sell Your Piano |

Advertise on Piano World
| Subscribe | Piano World | PianoSupplies.com | Advertise on Piano World | Donate | Link to Us | Classifieds |
| |Contact | Privacy | Legal | About Us | Site Map | Free Newsletter | Press Room |


copyright 1997 - 2014 Piano World ® all rights reserved
No part of this site may be reproduced without prior written permission