Why do some virtual instruments sound more realistic than others?

Question

This may be somewhat subjective, though it may be verified by a breakdown of the audio. But it seems easy to find like drums and other percussion, bass guitar, piano, harpsichord, maybe harps and organs, and others that sound realistic. But then like many horns or string instruments don't sound realistic. What is the reason for this?

Answer 1

Compressing 30 years of research into a couple of short paragraphs…

Non-percussive instruments are what's known in the trade as "an absolute b**ch to sample" because they can change over time, unlike a percussive instrument which has a relatively simple release phase after the initial impulse.

Some huge sample libraries use many many variations of each possible note & how it might be played. The user has to choose individual 'articulations' to persuade the ear that this is a natural progression. This is often combined with a 'round robin' structure which will play a different sample of each note each time it's re-triggered, to prevent the 'machine gun' effect of every one being exactly the same.

The player can help this along by modulating volume/vibrato etc over time, though the sound will not change to fully express the actual sound of a violinist playing harder or adding ever-changing vibrato.

There's a niche manufacturer who took an entirely different approach. Rather than each note being one of a selection of pre-recorded samples, each instrument is built from tiny snippets of sound, which can morph one into the other whilst the note is playing. This then sits within a mathematical model of the physics of the instrument string/wind/body.
Here's a demo - AudioModeling, SWAM

One thing to note - the harder it is to make each of these, the more expensive it is ;)

Answer 2

Any instrument that is percussive (you play the note and then there's almost nothing more you can do with it except stop it) is fairly easy to synthesize.

Any instrument where the player forms the note during the whole time it's played (and maybe connects it to the next note) are much harder to recreate. It's possible to make notes more expressive by modulating various parameters of the sound (volume, pitch, timbre etc.) but it takes a lot more work.

Answer 3

Others have mentioned more general answers, I'll try to provide a comparison of two real instruments.

You mentioned it was easier to find guitar samples, then it was to find stringed instruments (e.g. cello).

If we look at a solid body electric guitar, it's plucked and then the strings movement is measured by electrical signal through the pickups:

Although other parts of the guitar might also be vibrating, most the main part of the sound is coming from the pickups.

Therefore if wanted to build a guitar synth, we would have to model how the strings vibrates and just use that signal to create our audio.

Now instead if we have a bowed cello (which looks like this, and doesn't use a pickup)

we can see that sound is produced by the bow moving across the string (which is more complex than just plucking the string), which then vibrates a certain string, which then vibrates the entire body of the cello producing a rich tone which doesn't take any amplification to hear.

The rich tone also arises from chladni figures, which describe the way a thin plane vibrates at different frequencies so when you play a note, the front and back of the cello may vibrate in different ways:

(the white and black says how much each section vibrates)

So we can see that if we put the same effort into building a guitar synth as into a cello synth, we'll probably end up with a better sounding guitar synth as we'd need more effort to model the complexity arising from the cello.

So on average this is part of why string synths may not sound as realistic as guitar synths.