Comment by pierrec
3 hours ago
In this area of research, there's this classic trap that many fall into (including myself, many times). You focus on modeling things like the vibrating string, the resonant body, etc. to perfection. But it still sounds, uh, not great, because the more important and difficult part is modeling detailed control and the human/instrument interface. Air fluctuations around the violin sound like a fun experiment, but I don't think you'll get much additional realism from that, compared to a simple/classical impulse response model.
Even in this case, they're choosing the easy path (plucked, pizzicato), but the human/instrument interface is still audibly oversimplified while the resonant body has an unnecessary amount of "realism". The sound of pizzicato has a distinct character because the player's finger/skin slides a bit on the string as they're plucking, among other factors, which sounds like it's missing here. This can be tricky to implement because it's not necessarily a one-way impulse. The string is already vibrating and affects the finger, hence "interface".
This applies 10x more with bowed strings.
The point of this research doesn't seem to be to generate a nice sounding digital instrument, but to give violin makers a rough idea of what the instrument will sound like for different shapes / materials. This is useful for comparing designs, even if you don't simulate a human performance with complete fidelity.
So I don't know if your criticism makes much sense.
Either way, the demo doesn't sound like a violin yet. It's really cool that they got it from just a CT scan, and I get the usefulness if they're able to fill in the gaps.
Put differently, it takes years of practice to get a decent sound out of a real violin.
If your model doesn't sound like someone's strangling a cat then it's probably not realistic.