Jump to content
Maestronet Forums

Artificial Intelligence for Understanding Violins in The Strad


Michael_Molnar
 Share

Recommended Posts

I have argued that the next step in understanding violin acoustics lies in Artificial Intelligence (AI) which uses neural networks to make predictions. A few months ago on a trip to California I discussed this idea with a Silicon Valley computer expert, Luis Trabb Pardo. He opined that the issue was not just simulating the violin but more so the listener.

Well, as it happens, the September issue  of The Strad has a timely article, Intelligent Design by Sebastian Gonzalez. The last paragraph rings loud and clear with this sentence, "So understanding what’s so special about Stradivari’s violins is not something that can be accomplished without studying the biological response music produces in human bodies."

The article explains a first step in simulating violins using neural networks. I think we can agree that simulating the entire violin, not just components such as the free plates, must be the objective. 

Link to comment
Share on other sites

59 minutes ago, Michael_Molnar said:

"So understanding what’s so special about Stradivari’s violins is not something that can be accomplished without studying the biological response music produces in human bodies."

The study needs to be expanded to look at the emotional response to the name Stradivari and what it does to one's perception of sound.

Or forget the whole thing.

I certainly agree that studying parts, like free plates, doesn't get anywhere near describing how the whole instrument behaves, but I'm not seeing from here what a neural net is going to do in helping understand how to construct better instruments (I'm single-minded that way... not very interested in academic exercises).

We have had modal analysis and spectrum response information for years now, with no fantastic revelations leading to fantastic instruments that I can see. 

Maybe I'm just in a foul mood... but trigger words like "special" "Stradivari" and a fancy new high-tech method of investigation... they tend to set me off.  I should get back to the shop now.

Edit:  I wrote the above before reading the article.  I didn't get back to the shop, but read the article.  I suppose if you're a big fan of matching signature mode frequencies, this might help a bit.  Maybe after 30 years and a billion dollars, you could say something about the higher frequencies.

Link to comment
Share on other sites

2 hours ago, Michael_Molnar said:

So understanding what’s so special about Stradivari’s violins is not something that can be accomplished without studying the biological response music produces in human bodies.

That sentence is unchallengeably correct but begs a hundred questions at a more fundamental level. I'd like to see if AI could learn a task that many humans find pretty easy, for example identifying the three notes of a triad simultaneously played on piano or by strings. Not so easy when you only have the mixed Fourier spectrum to look at. Looking at the biological responses of cochlear hair cells and neurons in the brain, we really don't have much idea how humans do it and AI, even if successful, probably wouldn't enlighten us.

The leap from that to such exquisite discriminations as between a plain violin and a "great" one is cosmological.

Link to comment
Share on other sites

2 hours ago, Don Noon said:

The study needs to be expanded to look at the emotional response to the name Stradivari and what it does to one's perception of sound.

Or forget the whole thing.

I certainly agree that studying parts, like free plates, doesn't get anywhere near describing how the whole instrument behaves, but I'm not seeing from here what a neural net is going to do in helping understand how to construct better instruments (I'm single-minded that way... not very interested in academic exercises).

We have had modal analysis and spectrum response information for years now, with no fantastic revelations leading to fantastic instruments that I can see. 

Maybe I'm just in a foul mood... but trigger words like "special" "Stradivari" and a fancy new high-tech method of investigation... they tend to set me off.  I should get back to the shop now.

Edit:  I wrote the above before reading the article.  I didn't get back to the shop, but read the article.  I suppose if you're a big fan of matching signature mode frequencies, this might help a bit.  Maybe after 30 years and a billion dollars, you could say something about the higher frequencies.

Ya, well I still think we need to do the psychological/emotional response research. I think if you had a room full of violins, say 20, and a room full of UN-SOPHISTICATED listeners , say 100 , who could give a rats ass about Strad, Not telling them anything other than to list the ones from top to bottom they liked and you could somehow force them to go through listening test's , and then repeat this over and over, I do not think you would find Strad coming out on top, I honestly think it would be 50/50 odds for every violin there, from 120$ outfit violins to the Strads

And furthermore, that having any foreknowledge of what the instrument is, or really even any level of sophistication in regard to classical violin music would pollute their ideology and effect their emotional response about it.

Thinking of it like a jury where the lawyer was excluding a certain "type" of person from the lineup, I feel to get to the bottom of it you would need to ask the people who would be least likely to know such things.

Link to comment
Share on other sites

1 hour ago, matesic said:

That sentence is unchallengeably correct but begs a hundred questions at a more fundamental level. I'd like to see if AI could learn a task that many humans find pretty easy, for example identifying the three notes of a triad simultaneously played on piano or by strings. Not so easy when you only have the mixed Fourier spectrum to look at. Looking at the biological responses of cochlear hair cells and neurons in the brain, we really don't have much idea how humans do it and AI, even if successful, probably wouldn't enlighten us.

The leap from that to such exquisite discriminations as between a plain violin and a "great" one is cosmological.

Right and again I think that all gets right down to levels of sophistication, but then that begs the question, can not an un-sophisticated listener enjoy a crappy sounding violin just as much if not more than a Strad, and even if 100 concert violinist's told them they were wrong, would that person be wrong for enjoying "that" violin and its sound more? This of course slams right into bluegress and "twang centric" tone which is a completely different, yet , not to be disregarded "thing" nor are "hybrids" {which is what I focus on} .

And I would NOT like to see AI be able to do any task a human could do, but I know it's coming and hopefully I'll be dead by then. Pandora will not be stuffed back in the box

Link to comment
Share on other sites

So, for me, AI needs to be defined here...just neural nets?

To the article....there are many simplifications. Remember, all models are wrong but some are useful.

It appears the inputs to the NN are just some of the geometry. Perhaps they need to increase that a lot, perhaps by starting with wood perimeters and wood treatments/varnish etc. Small changes can make big differences and NNs are famous for not providing the way.

Using statistical methods to analyze large amts of data could give some objective categorization of violins. The categories would be ?? Peoples 'response'?? - way to subjective. Some frequency response curves?? But data might allow you to build a NN on an objective basis. But what does it provide in terms of understanding? What are you trying to understand? The research mentioned is not about identifying good tone. I think ya'all know the parameters of good violins better than anyone. Maybe, someday, further statistical analysis of 'good' violins would provide more knowledge about good violins. Maybe.

Also, applying a biological model might provide A Model but probably not one that understands violins. This could be used to construct a NN. But to what end? Sampling the audience would only get their bias....pick what ever audience you want. 

A thing that happened in Natural Language Understanding that created a big breakthrough were statistical models created by some physicists that basically relied on frequency of words, combination of words etc. to create disambiguation. No grammar, no stemming, no understanding. Perhaps I over simplify.

So I think you could create a model, by any means, and it would tell you that a violin that resembles that model would be well appreciated by a segment of an audience(database). How you actually create the pieces ( the actual understanding) would still be lacking. 

Link to comment
Share on other sites

I noticed that all VSA tone awards (as far as I know) go to experienced craftsman, not analytical geeks.  Until these researchers start blowing away the competition, I'll go with learning from the pros and some trial-and-error testing of ideas.  I consider myself a bit of a technical geek... but whatever tonal success I have I'd say is 90% those other things.

Link to comment
Share on other sites

AI is a catchall term for any computer algorithm that mimics some aspect of human thinking. It need not be a neural network.

The work presented in the Strad article was covered previously on this forum.

Here is the thing with neural net algorithms: you have to "teach" the net by changing an input and telling it what it should predict. It then compares what it predicted to what it should have predicted and then adjusts the "weight" it places on its inputs. You have to do this for a lot of different inputs and be able to measure the actual "result" it should predict for each of these cases.

Technically, the authors used a neural net to develop a multi-dimensional curve fit of geometry parameters to measured modes. There is nothing really "intelligent" about this approach. It is more a computationally efficient algorithm for handling many inputs, or a small number of inputs with a range of values, to predict a limited number of results.

The real question is the following: Is there something one can quantitatively measure that is an indicator of superior violin performance, like tone, bow response and projection. From many threads on this forum, it does not seem like mode predictions will fulfill that requirement. 

Link to comment
Share on other sites

1 hour ago, Dr. Ludwig said:

A more detailed paper by the authors.

https://arxiv.org/pdf/2102.04254.pdf

So I could be more enlightened after digesting this paper fully.B)

That was discussed here some time ago I believe.

While I strongly believe AI and similar modern approaches have something to say here, the violin business is not the area where use of these technologies will generate enough extra profit so implementation will not be worth the investment unless some scientific grant covers it.

I wrote it few times in old threads that these technologies have already been used in similar areas for many years. I still remember my first encounter with this at a math conference some 15 years ago where professor of applied math/ IT spoke about their research for a private concern where thay applied these tecnologies / theories (black box analysis, fuzzy sets, NN, AI) to replace human control of production processes with success. The whole thing was based upon learning from real people in real environment. The main application involved pulp/ paper making where tha main problem was just like in violin making variability of raw materials which is mostly wood. The experienced workers learn over time how to adjust the processes of the line to get best end results and the research managed to use AI to replace most of them very reliably using arrays of sensors. AI can easily learn from such live environment if you can provide measurements of sensors and good quality feedback. There are commercial AI packages that can replace human working at computer (like accountant or such) by "looking over his shoulder" for some time learning what he does with documents...

I can see someone like Yamaha apply this to their production. With good measurements of all possible variables and good feedback it is nowhere near impossible (this would require judging of quality of instruments that is quite subjective but with experienced player at the QC still doable) I don't see much use in workshop of single makers with limited output, their old school brains can usually do good job.

I think the Gonzales et al. paper shows quite different approach as it is completely removed from reality and they are trying to emulate theoretical model (FEM etc...) with AI. 

Link to comment
Share on other sites

7 minutes ago, ctanzio said:

The real question is the following: Is there something one can quantitatively measure that is an indicator of superior violin performance, like tone, bow response and projection. From many threads on this forum, it does not seem like mode predictions will fulfill that requirement. 

I don't think you need to measure the performance quantitatively - If you could do that you would be half way down to discovering how the darned thing works without any AI. Good old statistics and few good players could pretty much grade the violins into 5 or 10 qualitative categories and use that. Much like AI in OCR, the input images are mapped onto a set of characters...

Link to comment
Share on other sites

The qualitative categories need to have some numerical scale associated with them, like bad=0, good=1, awesome=2, for a computer to generate a scaled output.

There is also the question of the inputs.

There would have to  be a standard way to test each violin that would give you confidence that you are capturing the parameters you think might indicate a violin's performance.

Moreover, you don't just play into a microphone and expect the computer to make some sort of sense of it. You would need to give it some way of assessing the sound capture, like breaking it down into a response spectrum and extracting stuff like response peaks, relative strength of the peaks, density of peaks per octave, and maybe a dozen other things I can think of off the top of my head that may or may not be relevant to the sound and playability of a violin.

AI that reproduces human adaptability to problems that occur during certain tasks doesn't  need a neural net. It is a bit of a different AI problem. Classic examples I worked on years ago from the infancy of AI is teaching a computer how to be a medical expert in blood diseases, or diagnose problems with a fleet of diesel locomotives. The computer programs can accept and organize knowledge from human experts in the form of data points (object, property, value), to assemble and traverse decision trees to guide further testing and assess possible treatments/repairs.

They are known as "heuristic" AI algorithms because, in a very real sense, the computers can learn by observing, like humans.

Modern versions of these programs now have sophisticated interfaces that let computers gather the data points and do further testing without human intervention, but they still require human experts to train them, or at least spell out clearly defined goals that are quantified in some way.

I once had an airplane flight instructor teach one of these heuristic learning programs how to help an inexperienced pilot to land a plane. After a couple of hours of "talking" to the computer using a natural language text interface, he was amazed at just how good the program became at talking someone into landing a plane in real time.

We never actually used the program to actually help someone to land a real plane, as it was just a proof-of-concept of the program. The point is one needs clearly defined inputs and goals, all quantifiable in someway, even if it is just probabilities.

Link to comment
Share on other sites

3 hours ago, ctanzio said:

The real question is the following: Is there something one can quantitatively measure that is an indicator of superior violin performance, like tone, bow response and projection.

Dunnwald attempted that decades ago, and as far as I know that (or a similar power band scheme) is about all that is used to try to objectively evaluate tone.  I kindof agree with the general buckets and relations, but I think too much is lost in trying to boil things down too far.  In the end, you run into the problem that EVERYTHING is personal preference, and can not be objectified very well.

Link to comment
Share on other sites

I remember talks with my father, a physicist, who always showed me the latest articles on violins he found. (He was the type of guy spending hours in the library of the research center to hunt for information in an age where the internet didn’t exist. He did it mostly for his own research but apparently once in a while he was in the mood to hunt for other stuff)


And then he always came to me showing the most recent ‘exciting’ findings about violin sound or in this case the AI approach. And always my question to him was ‘Any idea how to make this?’ He would then shrug his shoulders and say ‘no idea’

This leads me logically to the only conclusion that we need ‘procedure oriented research’. Change things on an instrument and see or rather hear what changed. I am trying to do my small contribution with the new concept (super light?) violin. 
——————————————

When reading ‘Il manoscritto di Giovanni Antonio Marchi’ it is full of observations what properties have what influence on the sound. And though it is not scientific, it reminds me at times on Don Noons analysis talking about  ‘low string response’ ‘mid range’ etc. in the language of the 18th century. Though Marchi wasn’t the peak of violin making (probably because Stainer was his hero) it gives a glimpse on ‘procedure oriented sound calibration’ in the workshops back then. (Or learning by trial and error about arching thicknesses and material) In any case the ‘science’ in those days was alchemy, and despite its obvious shortcomings it had a highly developed approach to compare observations on procedures with the outcoming results. That’s mostly what we need to make better instruments.
 

For the AI approach I would at least hope that it can figure out what happens if you change which thing on a simulated model. This won’t give a complete answer on how to build good sounding instruments but hopefully prevent to think in wrong directions or show where alterations simply don’t matter so much. In this context I wouldn’t be surprised if such a model would show that some traditional beliefs are by far less important than other things nobody ever thought of. 

Link to comment
Share on other sites

46 minutes ago, Don Noon said:

In the end, you run into the problem that EVERYTHING is personal preference, and can not be objectified very well.

Maybe the only thing which can be objectified is sound projection or carrying power. And as we know from instruments past and present this does not boil down to exactly the same overall sound characteristics.

Link to comment
Share on other sites

5 hours ago, Don Noon said:

I noticed that all VSA tone awards (as far as I know) go to experienced craftsman, not analytical geeks.  Until these researchers start blowing away the competition, I'll go with learning from the pros and some trial-and-error testing of ideas.  I consider myself a bit of a technical geek... but whatever tonal success I have I'd say is 90% those other things.

No dips, just wood chips ;)

Link to comment
Share on other sites

11 hours ago, ctanzio said:

The qualitative categories need to have some numerical scale associated with them, like bad=0, good=1, awesome=2, for a computer to generate a scaled output.

There is also the question of the inputs.

There would have to  be a standard way to test each violin that would give you confidence that you are capturing the parameters you think might indicate a violin's performance.

Moreover, you don't just play into a microphone and expect the computer to make some sort of sense of it. You would need to give it some way of assessing the sound capture, like breaking it down into a response spectrum and extracting stuff like response peaks, relative strength of the peaks, density of peaks per octave, and maybe a dozen other things I can think of off the top of my head that may or may not be relevant to the sound and playability of a violin.

AI that reproduces human adaptability to problems that occur during certain tasks doesn't  need a neural net. It is a bit of a different AI problem. Classic examples I worked on years ago from the infancy of AI is teaching a computer how to be a medical expert in blood diseases, or diagnose problems with a fleet of diesel locomotives. The computer programs can accept and organize knowledge from human experts in the form of data points (object, property, value), to assemble and traverse decision trees to guide further testing and assess possible treatments/repairs.

They are known as "heuristic" AI algorithms because, in a very real sense, the computers can learn by observing, like humans.

Modern versions of these programs now have sophisticated interfaces that let computers gather the data points and do further testing without human intervention, but they still require human experts to train them, or at least spell out clearly defined goals that are quantified in some way.

I once had an airplane flight instructor teach one of these heuristic learning programs how to help an inexperienced pilot to land a plane. After a couple of hours of "talking" to the computer using a natural language text interface, he was amazed at just how good the program became at talking someone into landing a plane in real time.

We never actually used the program to actually help someone to land a real plane, as it was just a proof-of-concept of the program. The point is one needs clearly defined inputs and goals, all quantifiable in someway, even if it is just probabilities.

Agree with you. I think the problem of objective evaluating violins is just natural thing and even the best players or makers will evaluate the same violin differently. Same as competition judging. But it is not important to have "objective" exact rules. Every maker evaluates his own instruments with his own ears and help of a good test player or two. The same would work for this purpose. Give each instrument mark for tone color, carrying power, clarity, etc, whatever you consider important and don't look into spectrum or modes as you're looking into black box no one can decode right now.

Link to comment
Share on other sites

AI doesn't exist yet!

For now it's a marketing thing for the BIG GROWTH (hacking), that is endless....

But it will and it's not what today's developers and other enthusiasts imagens. Neural networks is an embryo that our grand children will laugh at and hate us for when we are dead.

Violins are not first priority for AI...

 

Link to comment
Share on other sites

16 minutes ago, Peter K-G said:

Violins are not first priority for AI...

E.g. underwater sound is processed using any possible means to identify the sources in military and intelligence applications. There are acoustic mines, stuff imitating a given sub or surface vessel to lure the enemy. Or makle things explode at a decoy rather than a real vessel.

In human intelligence voice recognition is a big "above surface industry" for forensic work, intelligence or whatever. We are good at it with out brains and trained listening system.

I have a collegue who made her master thesis on an AI network for identifying birds. There are studies in bioacoustics where they try to figure out how mice, rats and other guinipigs thrive in the lab by their ultrasonic and sonic vocalisations. 

My point is that there probably are numerous aplications out there which can investgate the violin and the player using AI. The bioacoustics example is a commerically available program based on a PhD. 

I think it is easier to figure out and identify the player than the instrument. But I do belive it is possible also to identify the instrument at hand, maybe also across players.

Link to comment
Share on other sites

Anders,

You and I are about the same age. One of the things I admire is knowledge (hence My Msc in Technology Competence Management)

I'm truly impressed by your knowledge in acoustics.

You might be interested by the knowledge in my domain, 30+ years as a software architect?

Link to comment
Share on other sites

51 minutes ago, Peter K-G said:

You might be interested by the knowledge in my domain, 30+ years as a software architect?

I’m not. Do not need it, nor am i curious.. Our room acoustic program use graphics and originally a program language. Im pretty good at that, in steps. 

If we need something we will pay attention and be curious.

 

Edited by Anders Buen
Spell correction and an extra sentence.
Link to comment
Share on other sites

4 hours ago, Anders Buen said:

I’m not. Do not need it, nor am i curious.. Our room acoustic program use graphics and originally a program language. Im pretty good at that, in steps. 

If we need something we will pay attention and be curious.

 

:)

When can you start?

As a CEO, you would be the one I'm seaching for...

Link to comment
Share on other sites

48 minutes ago, Peter K-G said:

:)

When can you start?

As a CEO, you would be the one I'm seaching for...

I think we are as different as possibly can be, only the age and some interest in violin making is common, possibly some education. I would never ask anybody if they were interested in hearing my professional history. But many males I know tend to listen for a few seconds then they start talking about themselves and their own ideas. I think that may be a mild kind of personal disorder or simply persons who need somebody to listen to them, because they do not have enough support on that side in life. Or it may be a dominance thing. 

Link to comment
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

Loading...
 Share

  • Recently Browsing   0 members

    No registered users viewing this page.



×
×
  • Create New...