PDA

View Full Version : Musician Modelling



Markleford
10-09-2005, 10:45 AM
I personally believe that "Musician Modelling" is an important next step in recreating ensembles. Why is it that computer composers can have highly detailed samples, yet still create a "flat" performance? Much of it is due to the fact that the composer doesn't know much (if anything) about how to play the instrument in question. Without that expertise, he or she is truly just "playing the sample", and he's bound to the expression that is imprinted on the data.

Even if the composer knows how to play an instrument, and knows how to get the best from the samples, the process of playing each part over and over (and over!) for each member of a large ensemble (e.g. orchestra) is extremely time consuming! Moreover, current control schemes are woefully inefficient for capturing a realistic performance. With towards 10 controls shaping a single instrument performance in JABB, and only two hands and (a maybe a foot or two) to record with at one time, each part takes several passes!

Without specialized virtual controllers like EWIs to "broadband" this performance data across, you have no choice but to make this a step by step process by recording slider or knob or mouse motions. While I'd love to see more specialized controllers emerge, I think there's more of a hardware hurdle due to lack of demand for them. And even if these controllers were available, a composer would then have to master them all before he or she is in the position to "save time" when making a mock-up (let alone cost restriction).

So my take on the solution is to enhance "human playback" software systems to include the quirks of an individual. Some HP systems right now capture the randomness of performance, and that's great, but I'd like to extend it to "personalities".

For instance, we have a trombone player named "Tony". Tony likes to really lean on low notes, so his mf on the bottom end probably creeps up towards forte. He's also a little slow to start a cue when he's been resting for a while (lips get a bit "cold"!). For fast passages, though, he gets a bit excitable and rushes them a bit and gets a bit lazy on the articulation so that he's perhaps slurring a bit too much. His slide positioning tends to be a bit flat, too, but he adjusts fairly quickly when his ear picks up on it (on longer notes).

Other players in the section would have different quirks to their performance in slightly different ways. "Shane" can play louder. "Jim" has a cold so he runs out of breath quicker, long held notes petering out gradually. "Bill" is quick on the ramp to a crescendo. All of these performers together create a more varied, richer timbral and textural response.

You could use these personalities as presets, tweak them to taste, or randomize from scratch. Additionally, you could introduce the concept of a "section leader". Say that "Irving" is first chair. He has a moderate (50%) amount of sway over the other members of his section. As such, everyone else's performance patterns will be "morphed" to approximate Irving's playing style. In this manner, you can give a bias to the perfomance, while still allowing some weighted deltas on the part of other players.

Then comes the beautiful part: you become the conductor, with a gestural controller or MIDI data track, and each of the performers reacts to your direction on an individual basis (based on their parameters, influenced by section leaders), so the ensemble lives and breathes not as a machine, but as the real thing.

I think it would be very nice to be able to play back the same line of notation for each of these personalities, through the exact same sample set, and be able to determine who is playing according to what you hear in the performance and how it responds to your direction. After all, if you gave the same trumpet to Lee Morgan and Freddie Hubbard and told them to play the same line without embellishment, they'd still be discernable via their tonguing curves, vibrato shape and speed, etc. (well, *and* a brass player does exhibit a certain sound based on embouchure, but let's not get too picky! ;) )

Of course, having access to different instrument sample sets will only enhance this. Let "Lee" play an Olds Ambassador. Let "Freddie" play a Calicchio 3/9L. Whether these sounds are made by samples, additive synthesis, or physical modelling, it will only serve to further distinguish their "personalities".

And really, much of this Musician Modelling can be done today. We don't have to sit on our heels and wait for some technology breakthrough. It is low-hanging fruit, just waiting to be picked, but it *does* require a bit of cooperation between sample library and host/notation program developers. Will we see it happen?

The goal would be to allow composers to be composers, not experts in tweaking MIDI CC messages or virtuosos in a handful of virtual instruments. Convincing mockups should be as easy as dropping notes on a staff with efficient, minimal gestural notation. Building your ensemble should be as easy as selecting "musicians" and pairing them with instruments according to the preference of your ear. And as part of that, using a new sample library shouldn't make you have to learn a whole new control scheme, wasting your time to become an expert on the ephemera of yet another piece of software. I'd love to be able to buy a single instrument "off the shelf" just for the love of the sound ("one Buecher tenor sax, please!") instead of having to discount it because I'd have to learn a new control scheme.

I'd definitely like to be a part of making this happen.

- m

loogoo
10-09-2005, 01:49 PM
Gee, maybe we could somehow combine GPO with 'The Sims' - then we could actually watch them play as well as listen to them. Of course, they'll also want beautiful houses and the most expensive cars... :D

falcon1
10-09-2005, 02:00 PM
It would be really cool!

Personally I don't like much to have go through all this midi cc programming stuff. It would be cool I could just choose character to play my music, like you say - I would become the conductor not the programmer :D

Fabio
10-09-2005, 02:19 PM
I agree: your is not a crazy dream but a feasable project.

The amount of programming, the CPU load resulting, the arbitrary decisions related to style, are a huge work, that can terrify every business oriented software house.

I see for your dream two possible ways to become real:

- the start of an open source team of experienced but young programmers looking for fun and challenging to show his ability to find new jobs and contracts.

- a slow step by step evolution of new HP versions: may be we will have Finale 2013 with Human Players enhanced ensamble maker, with 3D virtual reality sensors to conduct the score in real time, generating with your hands undered of controllers...

In both the case, I think that GVMISA (Garritan Virtual Musical Instrument Shop Advanced) will be used to give an instrument to the Human Players.
It will run also on MAC OSY for Pentium7 in your Home Media Wall (a wall of your house providing computer, multimedia and communication tools, and house automation all in one)

Fabio
10-09-2005, 02:31 PM
Re-reading my previous message, I feel the need of a clear up: I'm NOT ONLY JOKING!

It's really a meaningful description of a very good and possible project.

The fear is the amount of time/work necessary to do it, and some new or more extended application of existing technologies.

wes37
10-09-2005, 04:06 PM
"Wes" always shows up late, plays flat, and mooches some of the pizza brought by "Tony".

...or is that TOO realistic?

Sepheritoh
10-09-2005, 04:19 PM
Musician moddeling is already well on it's way I think. Some instruments lend itself easier for it.

Band in a Box has done a fair bit to moddel musicians on Piano and guitar. It is mainly done by analysing phrasing, harmonies and emphasis from existing recordings.

Classical instruments like GPO's library other than solo instruments would be almost impossible, as it is more about orchestration and choice of instruments. I think many has tried to computerise orchestrations, including BIAB, but I have not yet found anything convincing enough.

squoze
10-09-2005, 05:18 PM
similar discussion in this thread (http://www.northernsounds.com/forum/showthread.php?t=23927&highlight=personell)

wst3ae
10-10-2005, 09:22 AM
I personally believe that "Musician Modelling" is an important next step in recreating ensembles. Why is it that computer composers can have highly detailed samples, yet still create a "flat" performance? Much of it is due to the fact that the composer doesn't know much (if anything) about how to play the instrument in question. Without that expertise, he or she is truly just "playing the sample", and he's bound to the expression that is imprinted on the data.- m

I think you are spot on in both your analysis of the problems, and at least one avenue towards a solution... but, I am not sure that is all of it.

GOS, GPO, and some of the really large/high-end libraries have made tremendous strides towards realism, and in the hands of someone who really knows how to use the tools, the results can be VERY realistic.

But there is still that element of perfection, or rather the lack of randomness that can show through, and I don't know it if is solved by musician modelling alone.

If the musician modelling takes into account the differences between different instruments as well as different players, and randomizes other attributes then perhaps it is the solution.

And I mean no disrespect, but this search for more realism goes back to (at least) the early 1970's... it is the holy grail for synthesis. Way back in the dark ages it was unrealizable, simply because of the available horsepower. Subtractive Synthesis was tool of choice and it has limitations! You could do FM, in a VERY limited and not too reproducible form on the ARP 2600 and some others, and if you built a large enough modular synth you could do some primitive additive synthesis as well.

Then there was C-Sound, and all the other synthesis tools that ran on rather large computers, and all in batch mode. They were cool, and they led to some wonderful research, but they were not at all easy to use.

One of the more important lessons I remember was that it was next to impossible to synthesize a single violin, for example, but we could synthesize an ensemble that had a passing resemblence for the real thing, at least with respect to everything else that was going on then<G>!

My first sampler was the Mirage, I don't remember how much memory it had, but it wasn't a lot, and they managed to jam an awful lot of usefulness into the box in spite of that. And it was a huge leap forward. Huge. I also remember hearing a demo of the Kurzweil 250 at an AES show back then, and being completely spell bound!

And of course that was just the beginning... and at first most of us were satisfied with the samples we had. But after a while we began to recognize that they weren't a replacement for live players. They simply were not real enough. So we learned tricks!

One of my favorites was combining samples, which were static because of the technology limitations, with synthesis, which wasn't all that realistic, but it was dynamic. For example, the Yamaha TX-81Z had the ability to play in monophonic mode using all eight voices in a round-robin schedule. The first note would play patch #1, the second note would play patch #2, etc. If you were to use a cello sample in the Mirage, and eight slightly different cello patches in the TX-81Z, the result was really quite animated. Maybe too animated<G>, but it beat the heck out of static samples.

But I ramble...

I think this is a very important area for research, and I know of at least three companies that have already made huge investments in that research, not to mention dozens of universities!!!

There are two obstacles presently - the first will go away, it is the resources available for implementation, and that increases daily. The second will be a bit more challenging. How do we recognize, let alone quantify those elements that make a musical performance interesting? And how do we allow for the fact that you and I may have different opinions on just what makes a performance interesting??

Still, I really enjoyed your post, and your descriptions. It is still more food for thought on a topic that is near and dear to probably everyone in this group!!!

Take care, and thanks for the great post!

Bill

FirmamentFX
10-10-2005, 11:44 AM
This has been done for CG characters in the animation world - the program's called "Massive". Each "agent" has a brain which is pre programmed with certain parameters (character, reaction to outside influences, in the case of a virtual musician a set of rules as to how the musician plays, what styles he/she favours etc), and at runtime the agents interact with each other to create the scene (or in this case the song :) ).

To get a true *ensemble* feel, any virtual musician would have to interact with other musicians (bass locking with drums etc etc), so it would be a massive (no pun intended) AI/neural network system.

You could eventually go the whole hog and have a virtual conductor for a virtual orchestra. The conductor would be set with simple rules as to how they interpret music, and then just hit play and see what happens.

A system based on AI would also have the benfit (or drawback, depending on your point of view) that no 2 performances would be the same...

Very interesting thread.

jc5
10-10-2005, 01:19 PM
An open source group project for such a thing would be interesting...

I cannot say how much I look forward to the possibility of one day being able to actually conduct the 'performance' in real time via some sort of sensory apparatus. It isn't crazy at all - basic directorial gestures and the method of beating time are fairly standard and could well be interpreted/understood by specially written software.

Fabio
10-10-2005, 01:59 PM
An open source group project for such a thing would be interesting...

I cannot say how much I look forward to the possibility of one day being able to actually conduct the 'performance' in real time via some sort of sensory apparatus. It isn't crazy at all - basic directorial gestures and the method of beating time are fairly standard and could well be interpreted/understood by specially written software.

Yes, I was not joking about that (being involved in 3D headtraking programs for ophtalmic optics).

I joke only about Finale 2013 & Mac OSY...all the rest is not available yet, but we all are already certain that it will happen quickly.

Nickie Fønshauge
10-10-2005, 03:34 PM
Finale 06 has a Tempo Tap feature, where you can tap the beat on a MIDI keyboard. It might just be possible to hook some sort of MIDI sensory thingy up instead of a keyboard.

Karl Garrett
10-10-2005, 08:00 PM
DP has a tap-tempo feature that allows one to tap the tempo and thus conduct the speed of a sequence, with any midi device. I use GPO with my classical guitar for live performance. My old Mac doesn’t have enough gusto to handle more than just a few strings or such with Altiverb, but I'm dreaming of the future where I could play with a complete orchestra. When the G5 laptop comes out, we just might be able to take the show on the road.

It is great fun, but you sure can't get an itchy foot or the piece goes down the drain. :D

Karl

Christopher Duncan
10-10-2005, 09:06 PM
Darn! From the title, I thought there were going to be stylish pictures here of the Robert Palmer girls... :rolleyes:

FirmamentFX
10-10-2005, 10:08 PM
DP has a tap-tempo feature that allows one to tap the tempo and thus conduct the speed of a sequence, with any midi device.

There are several systems around that do that for live performance already - the 2 best known being RMS's Sinfonia (http://www.rms.biz) and the Music Arts Mapper / M.O.S.T system (http://www.musicartech.com). Slightly less known is Cygnet from Swansos (http://www.swansos.com).

Sinfonia and Cygnet are tap tempo based playback for MIDI sequences, the MAT Mapper requires programming (it's a hardware based EPROM system) and basically maps any number of MIDI notes to a single key on up to 64 (I think) channels. Controller messages can also be programmed on key hit. It is a stunning system - if very old.

M

jc5
10-10-2005, 10:09 PM
Yes, I was not joking about that (being involved in 3D headtraking programs for ophtalmic optics).

I joke only about Finale 2013 & Mac OSY...all the rest is not available yet, but we all are already certain that it will happen quickly.

:) Yes, it won't be long now (at least one hopes!) for these things to emerge. Some of them are already in development in a number of different quarters..

dewdman42
10-11-2005, 02:57 AM
Markleford, I agree with you 100%. If composers can render an audio file by entering notes into a notation program and tweaking a bunch of parameters as available to add realism...then certainly a computer can be automated to do this based on player models. Its not at all an outrageous idea. the main problem I see though is that the actual sample libraries are changing so often with new and different capabilities...and different ways of handling things. For example, let's say there were some musician models that worked wonderfully with GPO. Then GPO advanced comes out with more controllers controlling more things in new innoovative ways.... hasta la vista to your old models... It will be difficult to stay on top of it. Not impossible, but difficult.

Ideally, I think it could start with some generalizations about what makes up a musician model.. Than perhaps it will be straigtforward to adapt those generalized models to each different sample library..some better than others. This where it can get out of control mighty fast though...because as other people have pointed out already, much much research has been done in this area already..but the problem as I see it is that the data they have compiled is WAY too complicated and all encompassing. For this to be successful someone would need to abstract some basic modeling attributes..perhaps just the things that can be feasibly recreated by midi controllers with GPO, for example...and make sure that the generalized models are specified in such a way that they can be expanded in the future as new technologies emerge.

Today we have MIDI which we all know is painful inadequate..but be that as it may...people do some fairly convincing work with it... so its actually capable of quite a bit...but I agree...even with tools like GPO..it often requires a lot of manual and tedious work to get the convincing scores done.

I'd like to say this could be handled by MFX plugins... ha ha..maybe that's a first step.

Markleford
10-15-2005, 09:25 AM
I'd like to say this could be handled by MFX plugins... ha ha..maybe that's a first step.Indeed it is. ;)

Thanks for the replies, folks. Good to hear people are getting educated that it's not just the size of the sample library that matters, but how easy to use and expressive it is! :)

I'll report later as things develop.

- m