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

Macy, you build a rod for your own back with your own assumptions on player precision. In this case, the idea of "velocity noise" of the player. It is predicated on the idea that the player cannot reliably determine the end result of attempts to play at a specific dynamic levels.

You actually have things quite backwards, what is important is that the player feels the sound that is produced accurately matches whatever movement was actually performed. It's not about what the player attempted to do before the action was started, it's about the physical reality of what happened. You see, when you play from note to note, you are continuously adjusting your dynamics in real time based on the bio feedback (sensations of effort) coming through your hands and what sound you hear. It's not about any single note, it's about how the ability to continuously adjust your dynamics and reach the desired effect. Having only 127 steps is problematic because the player will often want to very subtly adjust their dynamic range but find that they aren't able to impart precise enough control over it.

Lets consider the case of a slow diminuendo or crescendo. The player wants to be able to start at one level, and continuously adjust their touch over a specified period and finish at a certain dynamic level. I can tell you from experience that I always find this situation exposes the limitations of the 127 step quantisation. You state it as axiomatic that I will have too much "noise" in my technique to be able to control my dim/cresc to a degree that MIDI can't represent decently. I believe in that circumstance I have a smaller amount of "player noise" than you believe exists. I never feel like that limitation is present on an acoustic instrument. Furthermore, you think you can prove that by a crude test as to whether a player can hit a note 4 times and achieve the same velocity reading. That's not how musicians play. They hear the first note, then they compensate on the second note, then the 3rd note and so on. Hearing the minute difference is critical is being able to adjust properly. You need to hear it in order for it to be fed back into the bio-mechanics of the player. It needs a high degree of resolution. I can't prove it because of the limits of MIDI, but I believe I could perform very precise dynamics if the resolution were there. It would actually drastically reduce the "noise" factor in itself. You suggest that the level of player noise as being inherently in the person. I'm telling you it's actually a reflection of the poor resolution of MIDI. The player loses confidence in minute dynamic control because there is no reward for that effort. I believe the level of player velocity noise is far lower if the resolution is there. One becomes aware of the control one has if they experience it. I could try to prove it by recording something on a real piano and representing it as a waveform. You would see the amplitudes moving under conscious control in a far more precise way than your player noise theory suggests. I'm sure I could find thousands of classical piano recordings which would demonstrate extremely high player velocity resolution, and low noise.

It feeds into the other point I made about mixing etc because once you become aware of the steps being limiting, you will notice it everywhere. Anywhere you use MIDI, you will eventually find a situation where you wish it were more precise and graduated. That part of it isn't player noise related in terms of the innervation of one's own muscles on a keyboard, but it's part of a larger picture because many people edit their recordings via MIDI editors or they may program the whole thing and use a PC based sampled piano and not perform it at all. In that case, your argument about player noise goes right out the door. We aren't able to even program the precision we want to hear! I know that for a fact from my studio work.

This strikes me as similar to discussing the physics of speaker wire with "audiophiles". I don't understand how you can assert that you can play and adjust your key velocities in real time to a repeatable precision that exceeds by a factor of 2 or 4 MIDI quantization (suggested in your last posting), but you are unwilling to do a simple test to prove to yourself that you could even play within the precision of the current MIDI quantization because MIDI limits your confidence. That strikes me as reasoning similar to audiophiles that won't participate in double blind tests.

Play a rapid run of 8 notes, a simple scale with each note velocity exactly 1 midi step above the other. Or play the 8 notes each 2 MIDI steps larger than the previous note. What's limiting you from doing it with that precision repeatedly if you are sure you could do 1/2 or 1/4 MIDI steps if only those MIDI steps were smaller?

I don't think we are going to get any further with this discussion, so I'm going to respectfully (and I mean that sincerely) check out. I was an audiophile for years back in the 60's and 70's, designing my own amplifiers and speakers (my thesis was related to speaker design), and I was never able to close the gap between the "technical audiophiles" and the "subjective audiophiles". There's really no where to go when an audiophile asserts he can hear things that can't be measured or explained by physics or contradicts physics, and doesn't want to participate in tests to demonstrate his claims of hearing acuity. Nevertheless, some of those guys remain my good friends today. So in the same spirit, I'll check out now and go play my piano. Thanks for the interesting discussion and best wishes.

I never said I was unwilling to do a test. I don't know how it would be administered though. You also completely miss the point with your test. The quantisation levels of MIDI prevent musically meaningful gradations in dynamics. You are looking at numerical values, but that's not how musicians play. You sound more like an experimental physicist than a musician, which is why you design experiments to prove what you already believe.



Again, staggering failure to comprehend. No musician plays to realise digitised numerical values on a 0-127 scale. A musician plays with dynamics to contrast one note to the next. Each note is compared to the last, they aren't separate events. Resolution is very important to realise that goal. Also, my entire thesis is that rough quantisation prevents musical feedback, so targeting whole MIDI numbers is completely useless proposition designed to obscure my point and support yours with no actual scientific merit behind it. The problem with quantisation is that you can't tell how close to the edge of the next number you are. You could return a reading of 52,53,52,53, yet have played as precisely and somebody else who returned 53,53,53,53. That's what quantisation does, it misrepresents some data points more than others. Only a high resolution test would actually prove what people can do.

What would be useful is to see if there is a difference between a gradual crescendo on a standard MIDI device vs a high definition MIDI device. I am certain the high-def one will be better. How do I know? Because I've got a real piano and it responds far better to fine dynamics. I can do a genuine crescendo on my real piano. On my DP, it sounds rougher and less authentic. If you can suggest a way I can subject my hypothesis to testing, I'm all ears. I don't see how I can do it because I only have normal MIDI. The best I can think of is to record a real piano and analyse the waveform. Not sure what standards I'd use to numerically represent the velocities though.

Finally, your comparison of my argument to speaker cables is absurd and designed to offend. I'm not so much offended as I am disappointed at the breakdown of your logical thinking. I haven't shied away from testing, if a valid test were proposed. I even suggested a test myself. Your test is pseudo science. It's limited by its own flawed technology. The idea of testing a player's dynamic sensitivity with a low sensitivity device like MIDI is glaringly illogical. The instrument needs to surpass the capability of the performer to give meaningful information as to where the player's natural limit lies. I actually will do your test at some stage. If I manage to do what you say is impossible, how would I prove that I haven't forged the data though?

It has been an interesting discussion, to be sure. But you have actually made no effort to test whether what I'm saying might be possible. All you have done is state your opinion and designed a test which actually supports what I've said every bit as much as it supports your statements. You need to falsify my argument to actually persuade me. If you are going to paint yourself as the scientist and me as the speaker-cable peddling snake-oil salesman, you'd better make sure I can't falsify your claims.

In that response to my post, Macy, three times, as you see, you went back to 127 values as being sufficient, when I never claimed it was not (it was not something I was addressing at all), so I'm not sure you got my point. But your post was interesting nonetheless. ;-)

To get back to the actual point of my post, though, if you are claiming--as you appear to be--that MIDI volume/timbre for velocity x is unimportant if you can't reliably hit x on demand, then why do you think even 127 values are needed? Why not drop all the even number velocities and have only 63 steps, unless someone can reliably hit a value and not be so much as one unit off? Heck, I don't think I can reliably get within 5 of my intended value i.e. my attempt to generate 52 might easily yield a 47 or a 57. Does that mean that there is no reason to bother with values 48 through 56? In that case, 13 MIDI values would cover everything, right? Is it your position that 13 values is indeed enough, if you can't reliably hit numbers with the range I suggested? I think that that would lead to jumpy sounding expressivity... because even though I may not be able to get "that close" on demand, my gradual crescendo from below to above that range will certainly hit a velocity or two within that range. So I think it is clear that there is value to having those velocities, do you agree? If you do, how would you square that with the premise that a velocity that can not be duplicated on demand is extraneous? Or do you think that 13 values would actually be enough, if that's as close as someone could reliably get on demand?

OTOH, this issue is not necessarily a result of 127 being too few velocity levels; I think it could also be a flaw in how the keyboard is mapping your velocity to the various values it is capable of outputting.

Thanks anotherscott, you explained some things about the gradual crescendo thing that I didn't go into. And I really think you can't underestimate the effect on the player if the sound doesn't match precisely what was physically performed. The player faces a continuous dilemma of how to gradate the sound. It becomes more of a coping strategy rather than an innate sensitivity. Under the Macy's scenario, the best a player could hope for is to learn to play in a quantised way with 127 levels (thereby being able to reliably hit certain numerical values) If you've ever had a classical piano lesson with a fussy teacher, they ask for dynamic sensitivity an order of magnitude greater than that!

That compounds the problem, yes. And I have tried remapping the curve to try and get more control. No matter what I did, I found I was still bothered by the stepping I could hear. Not just when I played myself, but even in a sequenced crescendo. I often find myself wanting the values in between. Look, it's not the end of the world, but it does affect me. It's not just a question of the action itself being the factor. My AP has a silent system in it and a MIDI sensor array so I can use the same piano for everything if I wish. I think high end DPs will opt for more levels in the future. It's part of the natural progression towards realism (albeit glacial).

Would it be possible for us to hear the 'stepping'?

Good night Mrs. Robinson ... wherever you are.

Actually it's, Good night, Mrs. Calabash, wherever you are.

Good morning!

Yes, there may indeed be audible differences between each of the 127 steps - perhaps even VERY audible differences sometimes, say, due to chaos. (e.g, striking a string that is already in vibration, or playing a synth with a very wildly varying sound). I think what Macy is saying that he feels that 127 steps is simply enough (possibly more than enough) to capture the velocity signal from a person. The reason it is enough is that we are simply not able to strike the key with enough precision to warrant more steps, and we're simply reproducing our natural noise if we add more steps.

There's a problem though. If there can be a noticable difference between any two steps, is it fair to completely discard all the possible steps in between? Possibly not, and if it were an issue, the solution to that problem is simply to add randomness to the MIDI signal, so that the intermediate steps are in fact invoked as well. This process is called Dithering This would create a new velocity signal that uses higher precision to store the values.

Greg.

It might make things sound more interesting, but it doesn't address my concerns. I am actively challenging the idea that human beings can't exert more control than the 127 offers. I don't accept that the amount of human error is larger than these steps. ie. it doesn't capture the velocity signal of the player - certainly not all players. I think the folks that designed MIDI foreclosed on that too early. It was done at a time when computers were very slow and it made sense to not over tax them. I don't think they studied human perception terribly deeply when they concluded what we are capable of.

Dithering introduces randomness, but IMO that's not what's needed. That would not assist in the player connection problem. In fact it would worsen it because your velocity would be even less accurate to your physical effort than it is now, some of the time. That means you can't make continuous comparative judgments and adjustments. Dithering would only make it sound more interesting for the listener who doesn't have to perform the music.

I'd rather go with the higher resolution model. It will do what dithering does anyway. The players who can't perceive the difference won't be affected. The players who can will enjoy a greater sense of physical control over the sound. The MIDI programmer will be able to create more sensitive renderings of their musical. The mixdown engineer will have access to the in-between values that he sometimes wishes he had. Everybody wins! I don't favour a standard with 64,000 values, mind you. That creates complication we don't need. I think quadrupling the resolution would be the right amount.

That's fine, however the way to prove this is by doing the experiment that Macy asked you to do.

Greg.

I was going to quit posting on this subject because I truly felt like my discussion with Ando couldn't progress any further. But I think your comments indicate that you want to understand what I'm talking about. If you do understand what I am saying and want to argue about whether I'm right or wrong, that's great. We will both probably learn something, and I may be wrong or I may be overlooking something important. It's when someone doesn't understand what I'm saying, but wants to argue about it, there becomes no point in further discussion.

So I'll respond to your comment, but first let me restate what I am saying again so we don't start with some false assumptions.

The point I have been trying to make is that I don't believe a player can repeatedly play the same piece of music with the same velocities from one repetition of the piece to another. I know I can't do it, because I've tested myself. I can't even rapidly play a simple sequence of 5 ascending notes with 5 fingers and get the same velocities every time, no matter how hard I try. But I know there are much, much better piano players than me. So perhaps there are players that can do it for such a simple example. (It's extraordinarily easy in a minute or two to test one's self to find out if you can do that simple example or not.)

I tried the example once again on myself a few minutes ago. I did 12 trials of playing F,G,A,B,C in succession using 5 fingers. (The test took less time to do than typing this paragraph!) If I throw out the 2 worst trials the peak to peak velocity deviation in the remaining 10 trials was never less than 10 MIDI levels for a particular note, and never less than 9 MIDI levels in a single run of 5 notes. And those were the best cases!

But that simple example not withstanding, I can't imagine anyone physically playing exactly the same velocities within actual music repeatedly from one iteration to another. I think that assumption is so logical, and common sense, that if anyone wants to dispute it they would have to provide some evidence to show they can do it. And again, that isn't hard to do. Just record the same piece twice and do a note by note comparison of the MIDI values to prove you can do it or not. Again, I can't even do it for the trivial 5 note example above. But if you believe I'm wrong, and believe some piano players will play the same piece of music with exactly the same MIDI velocities every time, then you can dismiss the rest of my point. Otherwise, please continue.


Let me be clear about what I am NOT saying because it seems to be the key misunderstanding that is causing all kinds of trouble in understanding my point. I'm not saying a player tries to play certain MIDI velocities for certain notes. It makes ZERO difference that a player can't hit specific target velocities on demand. Of course no player is even thinking about MIDI velocities while they play. The only thing that is important to my thesis is that a player does not play the same piece with exactly the same MIDI velocities every time. There is a natural physical randomness to the actual MIDI velocities that are played, no matter how hard a player tries to play the piece exactly the same each time. That is all that matters to my point, the velocity randomness, not an inability to play any specific velocity on demand. Playing a specific velocity on demand has nothing to do anything. So I haven't suggested throwing out any MIDI steps, or doing anything else based on the inability to play any specific velocities on demand.

The velocity randomness is the same thing as a velocity noise component applied to whatever is played. If you play a piece 10 times, each individual note in the piece will randomly vary over some velocity range during those 10 repetitions. The playing of the piece will not be exactly the same each time, no matter how hard the player might try. It will sound a little different in timbre/loudness each time because of the velocity noise component. The larger the velocity noise component the more the actual sound of the piece will vary from one playing to another.

What does this have to do with MIDI quantization? MIDI quantization is another velocity noise component. Every velocity struck is rounded off to the nearest MIDI value. The timbre/loudness will be a little different than what it might have been if there were an infinite number of timbre/loudness steps. But there aren't. There are currently only 127 MIDI velocities, so there is MIDI velocity quantization noise added to whatever piece is played. It's another random velocity noise component.

So what happens when a player tries to play a piece of music? Regardless of what the player attempts to play there are 2 components of velocity noise added to the playing which varies the timbre/loudness of each note in a random way. There is the MIDI velocity quantization noise, and the player's random velocity noise that results from the player simply not being able to play with exactly the same physical velocities each time they play the piece. There are technical methods to analyze what happens when two noise components are added to a signal. It's more complex than one might expect, so I won't discuss it mathematically. But common sense should tell you that if one component of the noise is significantly bigger than the other, then making the smaller component of the noise even smaller doesn't improve things much at all. If the random timbre/loudness variations from the player's velocity noise are significantly bigger (louder/greater timbre variations) than the MIDI quantization noise effects on loudness and timbre, it doesn't do much good to keep making the MIDI quantization effects smaller and smaller. How can one argue that they need smaller MIDI quantization steps in order to properly adjust one's playing on a note by note basis, if whatever adjustment one tries to make for the next note is also subject to a player's random velocity component that may add or subtract several times the MIDI quantization velocity to the adjustment they are trying to make?

So if you have followed me this far, I'm not trying to prove that 127 MIDI steps is enough quantization for playing a digital piano realistically. I haven't predicted what the best possible player's velocity noise might be, but from my own testing of myself I suspect it is considerably greater than the current MIDI quantization. And I didn't state how much smaller the MIDI quantization should be than the player's velocity noise to make it insignificant. What I attempted to do was present a model of how to think about the MIDI quantization issue relative to playing a digital piano live. In my world, it's easier to think about such topics when you have a useful model that provides technical tools to analyze a situation, rather than just saying I need 2x or 4x or 100x better quantization because I know it will help me play better. My own experience, from testing myself and playing, and perhaps most importantly from capturing and analyzing the detailed audio output vs MIDI velocities of typical high-quality digital piano software, which is without any question a much bigger issue than MIDI quantization - which I mentioned just briefly earlier - 127 MIDI levels seems reasonable. That is no proof, it is just a opinion. And in previous posts I expressed it as "it seems to me that we reach an acceptable point with 127 levels", "it seems to me that 127 levels is sufficient", "it persuades me that simple lack of more than 127 level is not a significant issue". Those statements are just an opinion ("it seems to me"). I'm tried to present a model for how to think about the issue, not provide data to try to prove anything.

(Finally, I'm not talking about using MIDI to mix instruments or about editing recordings in post production or stuff like that. So please save those arguments for another discussion - there are many other issues there that are just as serious as MIDI quantization.)

Isn't this all a moot point? If pianos can only measure and generate 127 levels, who will produce a piano library that can respond to more than 127 levels? No one.

So the questions become:
- Do enough people want resolution greater than 1 part in 127?
OR
- Can marketing induce enough people to think they want it?

- If so, can a keyboard be produced to suit?
- If so, is there sufficient market to justify manufacturing such?

In fact, if you look at the actual sample's audio output with only 127 levels, which sometimes includes massive non-monotonic behavior, i.e. you increase the velocity by 1 step and the audio output level drops to the same level as a velocity 5 or 6 velocity steps lower, or simply massive non-linearity i.e. you increase the velocity by 1 step and the audio increases by the equivalent of 5 or 6 velocity steps, why would one think the manufacturers would do any better with twice the MIDI levels?

This recent part of the discussion misses the point entirely. If you are trying to play a demanding or emotive work, there is absolutely no desire on the part of the player to play with robotic precision, such that they can hit notes with identical velocity. What they want is feedback from the instrument. They want to feel that the response from the instrument matches their input as closely as possible. For the most advanced/sensitive musicians, I would suggest, 127 steps represents only a small fraction of their range of expressive possibilities, and therefore the feedback from an instrument limited in this way frustrates them. There is a reason that Roland went beyond this limitation in the V-Piano. It may also be true of the AvantGrand.

This is different from, say, those trying to produce (modern) dance music, where exact repetition is a desired quality. 127 steps is probably more than adequate for this genre.

Nailed it.

I will do that test and let you know the results. I have one caveat though: I don't believe it tests what Macy think it tests. My hypothesis is that the lack of player feedback prevents good player control. I have thought of a good way to get round this though: If I use my acoustic piano, which also has a MIDI sensor array in it, then I attempt to play a series of notes at the same dynamic, I can use the feedback from the acoustic mechanism and real strings to moderate my movements, but the MIDI array can out output values on the quantised 127 step scale. If what I am saying is correct, I should be able to get more consistent MIDI values when playing through my AP than if I do the same test with my DP, which can only feedback 127 step audio to me. How does that sound? It should directly compare the influence of continuous data vs quantised data.

I never said anyone SHOULD play with robotic precision, or hit the notes with identical velocity every time. I said I believe they CAN'T play with robotic precision and hit the notes with identical velocity every time. Are you deliberating twisting what I said, or is just not clear to you when you read what I wrote?

So, after you get the feedback (audio/tactile whatever) what are going to do with the feedback that requires you to have twice the current MIDI resolution? Are you going to take that feedback and play the next note 1/2 MIDI step louder than you would have before you got the feedback? If so, how do plan to do that if you can't reliably play the note without the extra 1/2 step? How do you add an extra 1/2 step of velocity unless you can also reliably play it without the extra 1/2 step of velocity?