![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Datasets used for AI learning
It is safe to say that I am at odds with Ai datasets that are composed of commercial &or historical music data(midi files). Ai as a prediction tool only really offers a next note choice difference from that which can be generated by algorithms. So in the pursuit of composing an original piece of music, why would I want to influence my composition directly with an existing composition? Second issue, say that I want to influence my original composition with the compositions of an existing artist or particular genre, of these songs how many of them are in the same key or scale? Third issue, I feel the whole way in which Ai is attempting to determine an original concept from these online collective datasets is wrong. When I have an idea for a song and go looking for like compositions to drive my song idea, I don't listen to every other song in the genre to get an idea or drive my creativity. I find that one song and my head expands on the possibilities of that one song. Ai choosing which note comes next from a collection of songs is just not the way a musician chooses the next note.
If I were to create a song in the style of Jimmy Buffet, why would I look to Iron Maiden's music for inspiration? Or vice versa. I've noticed that when using these commercial or online available MIDI datasets, the datasets are very general & have nothing to do with what you're trying to create. And in the worst-case scenario include songs by artist who have directly lifted composition from others. The goal isn't to copy but to absorb the vibe of the music you admire, aiming to craft something that fits within the same genre or captures elements of the original. These constructed datasets are just filled with all types of music that have no musical relationship to one another. In most cases I have found the available datasets to be just huge collections of all the midi files you could possibly get your hands on. Another collection for internet traffickers to obsess over. If your goal is to create a song that is a collage of all the available music, then I guess use it. If you goal is to parody a song of your or someone else, then I guess use it. If you are wanting to create a more refined song that is more genre specific, you will have to weed through the dataset you downloaded to find what builds on your theme and start removing input data from the dataset. So there really is no control with out of the box datasets. This is why I often say, "We need more musicians that are programmers or more programmers that are musicians.", a sentiment I've shared before in this blog. And this is why I construct my own datasets for Ai training.
I am going to now pull apart these datasets and question the reason why anyone would even want to use them at all for original AI thinking with regards to composing original music.
AI computers view all datasets as text. AI doesn't hear. AI experiences "Jimi Hendrix" as a text conversion of analyzed information. This leaves out a huge part of what the music of Jimi Hendrix is and as I think Jimi would say, "No AI you are not experienced!" AI never experiences the Audio. To grasp how huge this is when understanding music & how people like one song over the other...I'll quote Mr. Xenakis..."it is why art, music & literature can lead to realms that religion only occupies for people.". This aspect of music will not be so easily comprehended by AI now or anytime in the future. Music is on a cerebral level of thinking that exists apart from the analytical. The best composers of our day can sit down and attempt to write the next great pop hit, only to be defeated in the court of public opinion by an individual who sings over a drum machine they bought yesterday. It is the magic, mystery & beauty of the world's cultural music.
The pure data .mid files that we use here in the studio are 2-5,7-10-minute edits from hour+ long algorithm generated midi. Pure data generated .mid can be infinite, with the notes, chords and drum patterns constantly in a state of change. Like a garage band improvising all day to find the hook in a groove. Commercial .mid is as structured as the songs they are. What we had to figure out was how to get AI to predict changes in our song from all of these infinite variation jam sessions. We didn’t want to control the prediction, so we turned our attention to dataset themes. Developing themed datasets was done by auditioning midi tracks back-to-back and determining which tracks best complimented the song concept. There is no right answer or set procedure to dataset theme development only that in doing so influenced the final outcome the song.
When I started training my first in house Ai model for music prediction, I used datasets like clean_midi.tar to get started. In those beginnings I would start to notice in the predictions made, note and melody artifacts if you will that sounded like something I have heard or note selections that were out of key. Was the transformer model at fault? No. The fault was in the dataset. So, you would pick apart the dataset... and find yourself asking...what does Jimmy Buffet & Orbital have in common? Why is this contradiction in my dataset? These two respected artists create music in two different themes, two different musicality. Problem is they are both found in the clean Midi dataset. and in some of the online datasets the two music forms are already concatenated together. As long as AI continues to learn everything through text, we are going to have to be its ears. Organizing datasets into songs that include Jimmy Buffet and the like or songs by Orbital and the like was where I started with in constructing my own datasets. Rather than collide two styles, better to develop a completed style and then rerun the process to meld the song to another style. This quickly boiled down to not only a specific style but in the same key. Once the midi datasets were into themes in key did the model produce predictions that complimented the melody & chords that built parts and retained the concept to the developing song.
Employing artificial intelligence to create a fresh Beatles track by utilizing a dataset exclusively comprised of Beatles music of the same key will lead to faster success compared to utilizing a dataset containing all pop Beatles inspired songs. A more original path to creating a new song by the Beatles might be to predict a hundred plus Beatles songs from a dataset containing only Beatles music & use only those predicted songs for a new dataset. Then to predict a new Beatles type song from the new predicted Beatles dataset. When you think about it, it's actually how a musician creates a new song from the influence of others. It's a second step away from the original that yields a new original. Whereas that first step away is usually a more alternative version of the original. A composition predicted from a dataset comprised of previously predicted alternatives is an interesting procedure for composing an original like. It would be interesting if an AI model would eventually predict the original song from a prolonged series of datasets composed of predicted alternatives. Coming around full circle if will. Say if we, start off wanting to predict a new version of Here Comes the Sun; so we build a dataset of all the versions of Here Comes the Sun and predict a number of alternatives from this dataset. Build a new dataset of only those first round predicted alternatives & predict new alternatives from this new dataset. And again, and again, and again...etc Util somewhere down the process road out spits the original Here Comes the Sun. That would be quite the conundrum. What just happened???
I think we all can understand that there is a part of all music in every new song that is written. Music composition is a cultural genetic transition. You can hear Jimmy page in so much of the rock metal genre. All music has its audible influence. It's understood that once you take you original predicted score, fasten sounds to the different tracks, that you and your audience relate that your song sounds like. But it's the truth in composition that while your song may sound like something at its core its original. This is music.
111724 This blog entry on theme datasets is developing. more to come as I work this ridiculous rant out of my head.
Commercial dataset links
https://paperswithcode.com/datasets?mod=midi
https://www.classicalarchives.com/newca/#!/
https://bitmidi.com/
https://arxiv.org/abs/1810.12247
It is safe to say that I am at odds with Ai datasets that are composed of commercial &or historical music data(midi files). Ai as a prediction tool only really offers a next note choice difference from that which can be generated by algorithms. So in the pursuit of composing an original piece of music, why would I want to influence my composition directly with an existing composition? Second issue, say that I want to influence my original composition with the compositions of an existing artist or particular genre, of these songs how many of them are in the same key or scale? Third issue, I feel the whole way in which Ai is attempting to determine an original concept from these online collective datasets is wrong. When I have an idea for a song and go looking for like compositions to drive my song idea, I don't listen to every other song in the genre to get an idea or drive my creativity. I find that one song and my head expands on the possibilities of that one song. Ai choosing which note comes next from a collection of songs is just not the way a musician chooses the next note.
If I were to create a song in the style of Jimmy Buffet, why would I look to Iron Maiden's music for inspiration? Or vice versa. I've noticed that when using these commercial or online available MIDI datasets, the datasets are very general & have nothing to do with what you're trying to create. And in the worst-case scenario include songs by artist who have directly lifted composition from others. The goal isn't to copy but to absorb the vibe of the music you admire, aiming to craft something that fits within the same genre or captures elements of the original. These constructed datasets are just filled with all types of music that have no musical relationship to one another. In most cases I have found the available datasets to be just huge collections of all the midi files you could possibly get your hands on. Another collection for internet traffickers to obsess over. If your goal is to create a song that is a collage of all the available music, then I guess use it. If you goal is to parody a song of your or someone else, then I guess use it. If you are wanting to create a more refined song that is more genre specific, you will have to weed through the dataset you downloaded to find what builds on your theme and start removing input data from the dataset. So there really is no control with out of the box datasets. This is why I often say, "We need more musicians that are programmers or more programmers that are musicians.", a sentiment I've shared before in this blog. And this is why I construct my own datasets for Ai training.
I am going to now pull apart these datasets and question the reason why anyone would even want to use them at all for original AI thinking with regards to composing original music.
AI computers view all datasets as text. AI doesn't hear. AI experiences "Jimi Hendrix" as a text conversion of analyzed information. This leaves out a huge part of what the music of Jimi Hendrix is and as I think Jimi would say, "No AI you are not experienced!" AI never experiences the Audio. To grasp how huge this is when understanding music & how people like one song over the other...I'll quote Mr. Xenakis..."it is why art, music & literature can lead to realms that religion only occupies for people.". This aspect of music will not be so easily comprehended by AI now or anytime in the future. Music is on a cerebral level of thinking that exists apart from the analytical. The best composers of our day can sit down and attempt to write the next great pop hit, only to be defeated in the court of public opinion by an individual who sings over a drum machine they bought yesterday. It is the magic, mystery & beauty of the world's cultural music.
The pure data .mid files that we use here in the studio are 2-5,7-10-minute edits from hour+ long algorithm generated midi. Pure data generated .mid can be infinite, with the notes, chords and drum patterns constantly in a state of change. Like a garage band improvising all day to find the hook in a groove. Commercial .mid is as structured as the songs they are. What we had to figure out was how to get AI to predict changes in our song from all of these infinite variation jam sessions. We didn’t want to control the prediction, so we turned our attention to dataset themes. Developing themed datasets was done by auditioning midi tracks back-to-back and determining which tracks best complimented the song concept. There is no right answer or set procedure to dataset theme development only that in doing so influenced the final outcome the song.
When I started training my first in house Ai model for music prediction, I used datasets like clean_midi.tar to get started. In those beginnings I would start to notice in the predictions made, note and melody artifacts if you will that sounded like something I have heard or note selections that were out of key. Was the transformer model at fault? No. The fault was in the dataset. So, you would pick apart the dataset... and find yourself asking...what does Jimmy Buffet & Orbital have in common? Why is this contradiction in my dataset? These two respected artists create music in two different themes, two different musicality. Problem is they are both found in the clean Midi dataset. and in some of the online datasets the two music forms are already concatenated together. As long as AI continues to learn everything through text, we are going to have to be its ears. Organizing datasets into songs that include Jimmy Buffet and the like or songs by Orbital and the like was where I started with in constructing my own datasets. Rather than collide two styles, better to develop a completed style and then rerun the process to meld the song to another style. This quickly boiled down to not only a specific style but in the same key. Once the midi datasets were into themes in key did the model produce predictions that complimented the melody & chords that built parts and retained the concept to the developing song.
Employing artificial intelligence to create a fresh Beatles track by utilizing a dataset exclusively comprised of Beatles music of the same key will lead to faster success compared to utilizing a dataset containing all pop Beatles inspired songs. A more original path to creating a new song by the Beatles might be to predict a hundred plus Beatles songs from a dataset containing only Beatles music & use only those predicted songs for a new dataset. Then to predict a new Beatles type song from the new predicted Beatles dataset. When you think about it, it's actually how a musician creates a new song from the influence of others. It's a second step away from the original that yields a new original. Whereas that first step away is usually a more alternative version of the original. A composition predicted from a dataset comprised of previously predicted alternatives is an interesting procedure for composing an original like. It would be interesting if an AI model would eventually predict the original song from a prolonged series of datasets composed of predicted alternatives. Coming around full circle if will. Say if we, start off wanting to predict a new version of Here Comes the Sun; so we build a dataset of all the versions of Here Comes the Sun and predict a number of alternatives from this dataset. Build a new dataset of only those first round predicted alternatives & predict new alternatives from this new dataset. And again, and again, and again...etc Util somewhere down the process road out spits the original Here Comes the Sun. That would be quite the conundrum. What just happened???
I think we all can understand that there is a part of all music in every new song that is written. Music composition is a cultural genetic transition. You can hear Jimmy page in so much of the rock metal genre. All music has its audible influence. It's understood that once you take you original predicted score, fasten sounds to the different tracks, that you and your audience relate that your song sounds like. But it's the truth in composition that while your song may sound like something at its core its original. This is music.
111724 This blog entry on theme datasets is developing. more to come as I work this ridiculous rant out of my head.
Commercial dataset links
https://paperswithcode.com/datasets?mod=midi
https://www.classicalarchives.com/newca/#!/
https://bitmidi.com/
https://arxiv.org/abs/1810.12247
