Suggestion for Modartt : add attack samples for piano realism

Well, personally, I love the attack on Pianoteq’s pianos. I’ve always found it very pleasant—especially on the Steingraeber, but generally across all of them. It’s perhaps in the resonance and the tail of the sound where I sometimes notice something artificial, particularly in the high register. I’ll take that as a personal observation, though I seem to recall this being discussed on this forum before; in any case, it’s something that can be fixed with Pianoteq’s editing options, which are fabulous.

We’re back at the eternal debate over Pianoteq’s 'realism'. It’s a curious matter; I suspect many people equate realism with how much it sounds like sampled pianos (which use processed samples, so they aren't 'real' either) or acoustic piano recordings (also not real, for the same reason), rather than the actual experience of playing the kind of grand pianos that most of us have only ever seen in shop windows or concert halls.

On the other hand, I find it unlikely that Modartt would add the option you mentioned. It would be like admitting their model doesn't work—something that I, and I suspect many others, don't believe to be the case. (But, interestingly, the opposite does happen—that’s exactly what current digital pianos do when they add physical modeling for resonance, dampers, and so on.).

I've done some reading on this before and iirc, the general consensus is that "touch" is mostly in attack velocity - but concretely, how a given pianist chains velocities and adds character to them with micro- and macro-temporal variation (cf. organists faking dynamics with very short delays), pedal use, and preceding key release. These give enough variables to create a unique fingerprint. My personal theory is that pianoteq accounts for these factors better than other libraries.

There's also a fringe piece of research that suggests that different acceleration curves that different pianists might use to hit the same velocity cause some elastic compaction of the hammer felts, thus changing the effective hammer hardness. This gives theoretical grounding for isolated notes of the same velocity sounding different with different pianists on the same instrument. But this idea is not well established and, at best, a very slight effect. I find it interesting because this is a mechanical difference that couldn't be picked up by the sensors, so no library would be able to capture this effect with existing hardware.

But at the same time, you're right that there's lots of psychoacoustic trickery going on with this instrument. Seymour Bernstein says in one of his lessions "the piano is a series of illustions" and demonstrates an in-note crescendo by leaning into the keyboard, and I defy you not to hear a crescendo as he does that . So those coexist with the real, lab-measurable acoustic differences.

Actually top sampled VSTi do that.They use a technique called DFD (Direct From Disk) streaming that balances speed and efficiency and only the initial first milliseconds portion of the attack. The rest of the sample is streamed after in real time and starts from the moment you hit the key . Only 5-50KB are stored in memory. , this is usually a parameter in the settings , so the total can vary  between 50-500 MB of data preload. A VSTi such as VSL synchron with 100+ velocity layers requires about 500 MB of RAM for pre-loading .

I believe in version 6 (or some other old one) they did exactly that (use attack samples) until they learned how to effectively model them too

Interesting, when I looked into this a few years back the number of velocity layers was ~12 in these libraries (at least according to word on the street, my research stopped at that). So these are true raw samples and not interpolations? Thanks for clarifying!

Still, to me as a pianist this only solves part of the problem; it's not that I need so many velocity layers because I'm specifically targeting layer 37 in my playing (say) and don't want layer 38 - it's that I'm putting multiple attacks on the same note into relation with each other. And how the new attack interacts with the residual string activity from the previous stroke is almost as important as the velocity differential. So in my book this can't be solved with a large number of samples (because there's too many combinations), it needs a dynamic model.

...btw there was something that was bothering me with the pianoteq attacks, purely on an acoustic level. I would describe it as slightly dull/mealy compared to an acoustic sound. For me the solution was to increase the internal sample rate (the step function of the simulation), which sharpened the transients for me. It's not an obvious effect but it resolved the slight issue I was experiencing. Since then I wonder if that is what people who don't like the "pianoteq sound" are talking about.

I’d be OK with a toggle to allow higher CPU usage in exchange for far more realistic sound.

If one is running on a raspberry pi then they can use the standard model.

But on an M4 Pro for example, there’s a LOT of power available…