Piano World Home Page Forums Our Most Popular Forums Digital Pianos - Electronic Pianos - Synths & Keyboards Sampled VS Modelled Pianos ???

I think the visual analogy is useful here. In TV production, 256 gray-scale steps are the norm. Because that is way too coarse for cinematic fidelity, a new logarithmic scale is being introduced. A picture made up of an inadequate number of steps is still a recognizable representation of its subject matter, but strikes the viewer as less than real - although you wouldn't be able to pinpoint any specific inaccuracy.

In sound, I would hazard a guess that, while we may not be able to discern exactly what is missing when music reproduction is limited to 127 steps of timbre or volume, we will know that it is not sourced from a real (acoustic) instrument. I suspect that we are able to discern sound differences in the thousands of steps at least, even if we cannot locate any one error or missing piece of information.

Whether this has any bearing on velocity layers, I don't know for sure, but would suggest that if we really want true digital substitutes for acoustic instruments, we will need to go beyond the technologies designed over thirty years ago.

Just to be complete, I would like to note that even if people can tell the difference between one timbre among 127 and the adjacent one, it doesn't mean that people can tell the missing timbres between those 127 have been left out when hearing a piece.

I guess we could do an experiment where we took imperfect samples extreme and rendered a MIDI, then did the same using a version of the same product with fewer levels. I would be interested to see. Actually better would be to record and render a MIDI using the VAX77 and PianoTeq and then render it again using standard MIDI (from the same file, if possible) and see if we can tell (blindfolded) which is which.

I am confident that I can not perceive these issues because software pianos sound and play beautifully to me, even though they have only 13-20 timbre levels and 127 volume levels.

To be clear, I would love for MIDI to be replaced by HD MIDI or something with more volume levels, and I have no problem with sampled pianos including yet more timbre levels either. I'm superstitious that way.

I think what you are overlooking is that it is not enough that different velocities produce different loudness/timbre combinations. The difference in loudness/timbre must matter to the listener or the player. If they don't matter, then they aren't necessary to the design of the digital piano. Take the simplest case of the same two notes played in succession on an acoustic piano. Suppose my intention if to play them as the same sounds. But I'm physically unable to always play them at the same velocity. Sometimes they are randomly the same but sometimes not. Sometimes one is slightly louder with a slightly different timbre than other. But they are always so close that I don't care, and the listener doesn't care if they are exactly the same. So what difference does it make whether those two velocities are replaced by a single velocity on the digital piano? No one cared on the acoustic piano whether they were exactly the same or not within the players ability to produce them. Only when the difference in loudness/timbre becomes significant enough because of the players ability, that the player or listener cared that they were different on the acoustic piano, is it necessary for them to be different on a digital piano. It seems to me that we reach an acceptable point with 127 levels.


Nope, the logic doesn't work. Sure you can hear a jump between 50 and 55, and perhaps between 51 and 52, when played as 50 and 55, or when played as 51 and 52. But you can't hear the difference if I tell you I will play 50 and 55, but really play 50 twice instead. The ability to control what is played is fundamental to the listeners ability to discern and characterize differences.

Perhaps you are hung up on the idea that an acoustic piano can play more volume/timbre variations than a digital piano? Of course it can. Will there therefore be a larger number of distinct random volume/timbre combinations played because of the player's inability to play exactly the same velocities repeatedly? Sure, because the MIDI standard limits each note (ignoring sympathetic resonances, etc) to 127 unique loudness/timbre combinations (regardless of whether they are created by 127 separate samples or loudness/timbre interpolation), whereas an acoustic piano has an analog velocity input. But the key word above is RANDOM. The player has limited control over the actual velocities he plays. At some point, his ability to control the repeatability of the input velocity reaches a velocity resolution limit, and a random velocity component is added to his input. That random velocity can be considered the input velocity noise. So each performance will be measurably different based on the velocity noise, whether playing an acoustic piano or a digital piano. That velocity noise will excite more loudness/timbre combinations on the acoustic piano because it is continuously varying analog velocity noise, whereas it will become digitized velocity noise on the digital piano.

But that simply means each performance of the same piece will be RANDOMLY different than every other performance even though it is otherwise performed perfectly (ignoring all other playing variables than velocity noise). Now the payoff question? Who cares? Does the player feel betrayed by his acoustic piano because it produces even more randomness to his performance than the digital piano? Can anyone even detect these random performance variations when he plays the same piece 100 times otherwise perfectly? Once again, if the result of this randomness doesn't matter on the acoustic piano, it doesn't matter on the digital piano either.

What you seem to be missing, or ignoring, is that while the resolution of the digital piano, in volume and timbre, needs to be better than the players ability to reproducibly play with precision, it only needs to be better, not infinitely better.

I'm an engineer, that amongst other things, spent many years designing analog to digital converters, so let me try describing this in a slightly more technical sense. If an application requires data at a particular noise level, then the A/D converter is designed to have enough effective bits of resolution to provide that noise level. Once the effective bit resolution of the A/D converter significantly exceeds the noise level of the input signal, the input noise is limiting and there is no need to design a better, more expensive A/D converter. A digital piano, whether it is modeled or sampled, is basically an analog to digital conversion of an acoustic piano. In this case, the repeatability of the velocity produced by the player is the input noise level to the A/D converter.

Once the piano's resolution significantly exceeds the players ability to reproducibly play something, then the natural variations in his ability to reproducibly play (the "input noise" in a technical sense) dominates, and there is no longer any benefit to increasing the resolution of the instrument. In this case, the noise level that determines the necessary resolution of this A/D converter is the ability of human to reproducibly play music at a performance level that the performer (or others) no longer find the variations significant. That's when you can stop adding more velocity resolution. It seems to me that 127 levels is sufficient.

I've written my own software to analyze digital pianos. It captures time domain waveforms for all 88 keys at all 127 levels, and plots the audio output of each note vs key velocity, and the audio output vs each note at a specified key velocity in a number of different ways. It produces various kinds of time domain and frequency analysis for studying the layer interpolation and layer transitions. I've studied a number of pianos, and I would sum it all up by saying you probably wouldn't intuitively believe how much you can't hear. In other words, the defects in the samples and/or their playing implementation is pretty poor in some cases, and has to rise to a considerable level before it becomes audibly significant (which it sometimes certainly does). But that persuades me that the simple lack of more than 127 levels is not a significant issue.

Macy, if a player can never access the spaces between the discrete steps, his playing will always be rounded to the next whole number. That means that the playing will always sound that little bit more robotic - the steps create a sameness. This, I think, is the nub of the problem. It is why you can't reproduce a truly acoustic experience with digitals as they stand - it is always a (somewhat crude) approximation.

Sameness - in the sense of monotony - is not what the sensitive player is after. A piano piece needs to live and breathe.

I rather think that the acoustic experience can't be reproduced authentically because of how the sound is delivered through the strings and soundboard as opposed to directional speakers and has very little to do with lacking velocity levels in DP's.

I think you're right about the sound delivery problem. But to me it's the totality of the experience. Advances need to continue to be made across the board (no pun intended). I don't think you can say we've achieved the holy grail in any one particular area, yet.

Not so. Professional digital video since Day 1 has been 10-bit based, with 876 steps from black to reference white. The steps have never been linear. They are a result of an optoelectronic transfer function that utilizes a 0.45 exponent power function. Consumer digital video is 8-bit based with 219 steps from black to reference white, and again is based on the same non-linear opto-electronic transfer function. It's actually quite difficult for the average person to discern the difference between 8-bit and 10-bit digital video, even using test patterns.

Correct me if I'm wrong, but I believe 10-bit has to be downgraded to 8-bit for delivery at the moment.

My apologies if I got the wrong gray scale number - I was going from memory.

The problem is, that as video and cinema become ever closer, the engineers are having to develop more sophisticated scales for increasing the apparent color resolution. If all were rosy, and 8-bit was fine, there would have been no need for 10-bit - or these new advances.

For those of us who DON'T suffer synesthesia, please explain how visual acuity relates to auditory acuity.

Wow, Macy, way to get technical. Time to start up your own DPBSD. Well, we can always use another dewster here on the forums (I hope neither of you mind the comparison).

So suppose a player on an acoustic piano hits velocity 50.0 the first time they play a piece, and 50.6 the next time they play the same piece, and those get captured to MIDI using a silent acoustic piano or Disklavier-like piano, but they are rounded off to 50 and 51 respectively in the MIDI files.

If you play back the second MIDI file with the rounded value on a Yamaha acoustic piano with a Disklavier mechanical system, do you expect it to make the Yamaha acoustic piano sound all digital, while the first file sounds acoustic on the same piano?

If you mean converted to 8-bit for delivery to consumers, then yes. That's what I said. Consumer video is 8-bit. But professional digital video is 10-bit, and film-to-video conversion (telecine) is normally now captured as 10-bit.


The main reason for 10-bit video is that it allows digital processing to be done to the images with less degradation, rather than because 10-bit is noticeably better than 8-bit for viewing purposes. In fact, 12-bits can also be used for the same reason. Digital video projection in commercial theaters uses somewhat different technical standards than consumer video sources, so the film-transfers have to undergo additional processing for that and other reasons. (Digital cinema performance standards are actually less demanding than the performance of the best front-projection home theater systems.)

Sorry, I normally try to steer away from getting technical. But some ideas are just easier to express (not necessarily easier to understand) if you don't limit yourself.

Dewster seems to be doing just fine without any help. I'd never have time to play piano (or my other interests) if I tried to do what he does on the forum.

All I can say is "wow" to how much time people are investing in this topic. How about some practicing.

During the day my toddler won't let me play without sitting on my lap to pound some notes out herself. Lately at night my youngest has been sleeping in my piano room. Mother's out of town so she can't take them away. So no practicing for me this weekend.

But I can forum all you want!

Macy, I can't agree with some of your arguments. The velocity resolution is very relevant if you have ever tried to do complex mix-downs. When you are trying to get different instruments to sit in the mix so that they are prominent enough and balanced, sometimes you can spend a great deal of time editing MIDI velocities up or down by one value. I have literally had the problem where 52 is not enough and 53 is too much. Yes, there is a subjective judgement on my behalf on what sounds right (others may not care about what I care about), but the fact that I'm frustrated by the lack of values in between the 127 quantised steps is enough proof for me to say that the current MIDI standard is not up to the task of high definition applications. Put simply: I can hear the quantisations.

By contrast, whenever I've been in an analog studio, I've never felt constrained in finding the right amount of adjustability (obviously only talking volume in that situation). Whether you are talking volume or timbre, the problem is the same. People are buying into an assumption that people can't hear or feel the difference in levels, but these ideas haven't been tested thoroughly. Therefore it's just as valid for me to argue that MIDI needs more resolution as it is for any of you cynics to argue that humans can't hear or feel the difference.

I predict that in the future sometime, this problem will be addressed. I don't expect the wider audio world to adopt it, but if you are putting all this effort into trying to emulate a fine instrument like the piano, sooner or later the bottleneck has to be addressed. Resolution is a clear bottleneck.

Regarding another point you made: randomness is a part of music. DPs reduce that variable. It leads to listener fatigue. There is a sameness and flatness about the way DPs respond. four notes of 52, 52, 52, 52 is not as interesting as four notes slightly either side of 52. I have no doubt that the quantisation is responsible for that lack of interest factor. Even if I played a single line melody with no chords, to leave out the sympathetic vibrations element, there is a sameness about the tone that you don't get on an acoustic piano. I can play an acoustic piano all afternoon and not tire of its sound. DPs leave me cold after a short time. They don't respond dynamically enough for me. They don't even encourage you as a player to be more dynamic because you won't be rewarded for the effort.

You make some complex sounding statements of a mathematical nature, but none of it addresses the human perception aspect. Unless you are prepared to model human physiology and express it in terms of bits and resolution, none of your arguments about what human can perceive are persuasive - simply because you start with untested axioms about what can be perceived and how many bits are needed to "fool" the listener.

The inventors of audio compression, are a perfect example. They used scant data to decide which frequencies could be perceived directionally by humans. They based their compression ratios on assumptions about the limits of human perception. Then later, it turns out that many people can and do hear how "lossy" some of these formats are. I think some of your arguments are somewhat in the same direction. There are a lot of factors, not just hearing, which need to be considered. As a player, there is a very dynamic feedback between muscular effort and sound. They need to be highly concordant with each other. If there is any kind of mismatch between sound and physical effort/feedback from the action, the result is a less than complete experience. 127 levels is manifestly inadequate for me. It frustrates me on a regular basis in the work I do. I'd like to see MIDI at least double in resolution. Preferably quadruple. I think I would be happy with 51.0, 51.25, 51.5, 51.75, 52.0. But 51, 52, 53 - that's not even close for my purposes.

I suspect that anything replacing MIDI will have thousands or millions of levels and a much higher possible resolution in the temporal direction. It's just going to be quite an effort to get people to move over, especially without a clear, unified standard to move to. Professionals are loath to change. MIDI came out in, what, the '80s? And still going strong.

I am not aware of 127 being an issue in my playing, but I can easily imagine cases where it would be limiting.

I am in total agreement with ando - I think he expressed it much better and more fully than me.

No problem but lets examine what you think I said that you are now disagreeing about.


I never made the argument that you can't hear the MIDI quantization. In fact I said in the very beginning "Sure you can hear a jump between 50 and 55, and perhaps between 51 and 52". So thanks for calling me a cynic. I don't mind people disagreeing with me when they actually disagree with something I said. But since I never said that, you are arguing with your own misrepresentation of what I said, and I do object to that.

The issue that I responded to had nothing to do with whether or not one can hear the quantization step size. The issue I responded to was whether the quantization step difference in volume/timbre was significant compared to the effect of velocity noise caused by the limit of a players ability to repeatedly produce the same exact velocities when playing a piece.

And of course only that single variable was being discussed, not additional differential or integral non-linearity issues, or non-monotonicity - which can be present in sampled or modeled pianos. I mention those issues because they are very relevant to the particular application you are talking about. For instance, on a specific piano you might raise the MIDI velocity by 1 step on a single note at a layer transition, and it may have the same relative effect as raising or lowering the note amplitude by half dozen or more MIDI steps within a layer. Hence, when you think you are making a slight 1 step correction it might be many times larger than expected (and even in the wrong direction) when made near a layer boundary. I know that because I have made those measurements many times. They are not caused by the quantization step size but rather by poor product development.

Anyway, none of my remarks were directed toward using MIDI to mix instruments, or to fix a recorded performance in post. All of my remarks were quite clearly, I believe, directed to real-time live piano playing. Using MIDI to mix instruments, rather than using audio, and the subsequent change in timbre associated with changing MIDI values when you do that, is another issue. I'm sure there are many MIDI extensions beyond quantization that could be made for that purpose.



Here we do disagree. The velocity noise while playing is considerably greater than the MIDI quantization size. You don't have to believe me. Just try it. Record a piece, and then record it again. Compare the MIDI values note by note between the recordings. Are all the note velocities between the two recordings the same or within 1 MIDI step of each other? If not, the problem you speak of doesn't exist, at least not for the reason that the MIDI quantization is too large. (It can easily exist for other reasons, not enough sampling layers, not enough dynamic range, etc.)


Now you are completely misrepresenting what I said. I didn't make any mathematical statements at all. Nor did I make any argument at all about what humans can hear or not hear. I didn't start with any axioms about what can be perceived or how many bits are needed to fool anyone. That's totally wrong.

I said that it seems to me that 127 levels is sufficient, and not a significant issue, because the random velocity noise of the player is greater than the MIDI quantization size. That's it. It has nothing to do with what quantization levels you can hear in audio. If you don't understand that, then you don't understand what I was saying at all.


Well, I can't play the same piece of music twice and get the same MIDI values for every corresponding pair of notes in the two recordings, but if you can, then I would agree you need more than 127 MIDI levels.

with a big smile on my face ... +1