Any artifical intelligence/machine learning folks in here?

[nonohmic]

I don't know how to build a neural network, otherwise I'd try this myself.

How about using machine learning to analyze fretless players and assess their intonation?

Perhaps you'd need to isolate bass tracks first. Then you could either input the transcription as labeled data, and see how far off the notes are.

Training data sets could be simple tunes that either or real world played, or put slightly off pitch beforehand.

Or you could used unsupervised learning and just let the machine analyze the track, assigning the notes as it sees fit and then looking at it after. Presumably the further the machines output bass line is away from the transcription the worse the intonation??

Then we could answer once and for all who has the best intonation!

Someone tell me this isn't crazy..

 

[Michedelic]

Where’s the fun in all that?

 

[Oddly]

Seems like overkill.
Couldn't you just use a tuner?

 

[Selta]

I don't know how to build a neural network, otherwise I'd try this myself.

How about using machine learning to analyze fretless players and assess their intonation?

Perhaps you'd need to isolate bass tracks first. Then you could either input the transcription as labeled data, and see how far off the notes are.

Training data sets could be simple tunes that either or real world played, or put slightly off pitch beforehand.

Or you could used unsupervised learning and just let the machine analyze the track, assigning the notes as it sees fit and then looking at it after. Presumably the further the machines output bass line is away from the transcription the worse the intonation??

Then we could answer once and for all who has the best intonation!

Someone tell me this isn't crazy..

This would be fun, though it'd be quite the undertaking as well right now. I don't do a lot of ML/AI at work (I'm a software dev), but have played around with it here and there. I could see an undergrand taking this on as a great portfolio piece - it'd be complex and involve quite a lot of challenges, some of which you've touched on, and many others (glissando, chordal tones, etc. - let alone the technical challenges).

 

[DirtDog]

Are you talking about using this real-time, when the player is playing…or to assess a previously recorded peice?

 

[Michedelic]

On further examination, I’ve come to the conclusion that my mother could do that sort of thing without all the R&D; you know, analyzing and criticizing. Will there be a logarithm for sitting up straight?

 

[nonohmic]

Seems like overkill.
Couldn't you just use a tuner?

I guess the while idea of AI ML is that we don't have to do anything. For better or worse.

 

[nonohmic]

This would be fun, though it'd be quite the undertaking as well right now. I don't do a lot of ML/AI at work (I'm a software dev), but have played around with it here and there. I could see an undergrand taking this on as a great portfolio piece - it'd be complex and involve quite a lot of challenges, some of which you've touched on, and many others (glissando, chordal tones, etc. - let alone the technical challenges).

Right, so many variables eh. Didn't consider glissando and slides! But yeah someone will tackle it some day I'm sure.

Ill eagerly await the Adam Neely video..

 

[nonohmic]

Are you talking about using this real-time, when the player is playing…or to assess a previously recorded peice?

Well in my experience a lot of this work is done on existing data sets (recorded) to begin with. With the ultimate goal of applying it in real time (live music).

If someone in AI ML really wanted to tackle this within a few years you could have a machine watching you practice and telling you how far your intonation is off at the end of the practice session.

Where’s the fun in all that?

Now where is the fun in all that, indeed.

 

[bholder]

I started walking down the AI path back in college decades ago, it got really hard really fast (and this from a Physics / Comp Sci / Math triple major). Everything's changed since then anyways.

 

[DirtDog]

Well in my experience a lot of this work is done on existing data sets (recorded) to begin with. With the ultimate goal of applying it in real time (live music).

If someone in AI ML really wanted to tackle this within a few years you could have a machine watching you practice and telling you how far your intonation is off at the end of the practice session.



Now where is the fun in all that, indeed.

OK, real-time makes more sense. If assessing a recorded piece, you don’t need AI, you just need Melodyne and an isolated bass track.

This still might be a rule-based problem in that in order to classify any note played as properly intonated, you’d just need the frequency info for the note and then compare that to the 12 critical frequencies for the 12 semitones per octave (and their related harmonics within three or four octaves).

what’s your use case? For a teacher to assess a student? For a player to assess themselves?

 

[Jack Deth]

Since 2017 I started toying around on an Octonion based neural net, but never have enough time to work on it. I’ve pretty much given up on all my tech related ideas because it depresses me that I can’t actually find any personal time to do any of the cool coding we developers love to do. Being in management now in I.T. sucks the life out of most personal things. It takes everything I have to stay into music and that leaves literally no other personal time. If I had the time, then I would dust it off and yes it would work for this use case.

 

[Jack Deth]

what’s your use case? For a teacher to assess a student? For a player to assess themselves?

I personally would use it to analyze my technique and identify trouble spots for me to practice on. Using software that simply tells you “this note is N cents flat/sharp” is ok, but a neural net could tell you “when you play an F major progression in 3/4 time the Bb is 15 cents flat 90% of the time, but in 4/4 time it is 10 cents sharp 50% of the time.” That would tell me exactly what I need to practice on and closely monitor my technique.

 

[Stewie]

Find someone with a good ear. Have them watch and listen to you playing. When they grimace, you missed the spot.

 

[ardgedee]

If someone in AI ML really wanted to tackle this within a few years you could have a machine watching you practice and telling you how far your intonation is off at the end of the practice session.

That doesn't sound like AI is called for. Record the track, log the recorded pitch and nearest true pitch for each note played, and the rest would be pretty straightforward statistical analysis.

 

[T-boneParish]

This is a surprising topic to see on TB!

I'll play the game. I do AI/ML and, as a matter of fact, I pulled up TalkBass this morning because I'm procrastinating from getting back into a coding project.

There are some things that machines can do better than humans. And there are some things that humans can do better than machines. For instance, it's very easy for a human to look at a picture and tell you what's in that picture, but machines require tons of training to be able to do the same thing, and even then they are not as good as humans.

If the goal is to develop an AI algorithm that ranks a bass player's technical accuracy then, yes, this is doable. But it would tend to reward players that play like machines. They are the ones that have the smallest variability relative to some "perfect" benchmark performance.

But music isn't about performing like a machine (most music, anyway). Most music is about creating a vibe that connects with listeners. And humans are a better judge of that connection.

I'll connect with players that have emotion and rock-n-roll-attitude over those that have technical accuracy any day of the week. And that is true, regardless of what some computer tells me is "better."

 

[nonohmic]

That doesn't sound like AI is called for. Record the track, log the recorded pitch and nearest true pitch for each note played, and the rest would be pretty straightforward statistical analysis.

Lol I hear this every time I present a problem I think is good for AI. I generate terabytes of microscopy data in my job and the AI ML folks always want to analyze it. They then ask me what I want to find out..

I say x

They say...that's not an AI ML problem

So I say, well what is an AI ML problem..

They say well we don't know the data.

I say ok can you help me with x

They say.. that's not an AI ML problem. And repeat ad nauseum. I'm slowly getting better understanding though.

Anyways if you want to sit down and analyze every one of the millions of notes Jaco and Gary Willis played be my guest. I'd rather train a machine. Esp for those passages that are so fast even the waveform is blurred..esp on old recordings.

 

[OldSchoolFlats]

AI programming takes a good knowledge of linear algebra, statistics, and calculus. Also programming languages such as Python.

On the other hand "pitch correction" is rather tried-and-true. Think of all those songs where the voice sounds synthetic: you can take someone with a 'tin ear', use pitch-correction (or whatever the term is), and give them perfect pitch.

You might google Cakewalk, and see if there are free plug-ins where you can record tunes to a track, then analyze their intonation. The biggest issue would probably be to get isolated bass tracks. In a studio, each part is on a separate track. With tunes you buy or stream, everything is on 2 tracks: left and right.

 

[logdrum]

I did AI/ML for 4 years in San Diego although I worked mainly on HW accelerators. It will be a lot of work to build your dataset.

Plus what tuning do you want to gauge against. Stretched or equal tuning. A play upright bowed as well as saxophone and flute and intonation is never mechanical and it depends on the piece or how the orchestra/band or the piano is tuned. Sometimes it depends on the key. That is the nature of viol, brass and woodwind. That is why a string quartet sans piano sounds more in tune.

Did you do any of the AI/ML courses, like the Stanford version. You have dropped some buzz keywords in your OP so I assume you have familiarity with the subject.

I think NVIDIA sells a development kit that has audio capabilities. My buddy has been asking me to help him with some voice projects and is supposed to send me this kit.

 

[nonohmic]

OK, real-time makes more sense. If assessing a recorded piece, you don’t need AI, you just need Melodyne and an isolated bass track.

This still might be a rule-based problem in that in order to classify any note played as properly intonated, you’d just need the frequency info for the note and then compare that to the 12 critical frequencies for the 12 semitones per octave (and their related harmonics within three or four octaves).

what’s your use case? For a teacher to assess a student? For a player to assess themselves?

I personally would use it to analyze my technique and identify trouble spots for me to practice on. Using software that simply tells you “this note is N cents flat/sharp” is ok, but a neural net could tell you “when you play an F major progression in 3/4 time the Bb is 15 cents flat 90% of the time, but in 4/4 time it is 10 cents sharp 50% of the time.” That would tell me exactly what I need to practice on and closely monitor my technique.

My use case was finally putting to rest once and for all the TB circle jerk about which famous player has the best intonation

But yes, as Jack said. If it tells me my higher notes in passages tend to be more off than my lower notes then I should work on my pinky dexterity higher up the board, for example.