Hi all,
Not a musician here, but an AI master student instead. I am currently doing my master project and I was hoping if we could help each other out.
To give you an idea what I am working on; it can be best described as a sort of dynamic low- and high-pass filter. The purpose of this system is to separate piano recordings by the playing hands. Simply put, an audio file of a piano recording comes in, two audio files come out. One contains the left hand performance, the other contains the right hand performance.
Considering that such a task (piano hand separation) has not been performed yet on raw audio, I wanted to keep it as feasible as possible by at first focusing on Boogie Woogie only. My assumption here is that Boogie Woogie would be the easiest genre to separate by playing hand due to the strong contrast between left- and right-hand performance that already exist.
The development of such systems does require data however, piano recordings in this case. I would need isolated recordings of left-hand and right-hand performance (playing Boogie Woogie). These are essential for evaluation purposes, but also for training a supervised Machine Learning approach (basically learning patterns from examples). As you might imagine however, such data does not exist. Therefore, a dataset has to be built.
This is where Wikiloops comes in!
So what would I exactly need? I am very aware that this a lot to ask, but I am aiming for at least around 50 recordings where the left- and right-hand are isolated in separate files. The reason being for why I would require so many recordings is mainly for training purposes. Supervised Machine Learning systems really do require many examples such to make a good generalization of the data it is training on. The timeline to finish this dataset can be easily spread out across 3 months however, as I can start off with a smaller dataset for prototyping and just scale up whenever the data is available.
I hope that there are member who are willing to lend out a helping hand. Not only out of sheer good-will however, because I want to propose something as well. Consider that the availability of good data is perhaps one of the greatest challenges in Machine Learning nowadays, this includes the field of music processing and blind audio source separation. The positive impact additional data has had on performance can for instance be seen in the MUSDB18 4-track music source separation task. MUSDB18 is a dataset of 150 songs where each song is separated in 4 tracks, namely: vocals, drums, bass and other. The purpose of this task therefore is to separate these tracks as good as possible. When looking at the best performing systems on this tasks (https://paperswithcode.com/sota/music-source-separation-on-musdb18), the top two were trained on additional data beyond the initial 150 tracks from the MUSDB18 dataset itself.
Now, consider that Wikiloops has roughly 151K backing tracks WITH EVERY INSTRUMENT SEPARATED IN ITS OWN TRACK, this website is in my opinion really a gold mine for AI music scientists to be discovered. If we manage to build this dataset, I will make sure to credit Wikiloops appropriately in any academic work I am going to publish.
I know that "promising exposure" is a laughed at cliche for artists nowadays, however the fact is that increased exposure of Wikiloops to the scientific community could lead to more traffic on Wikiloops and in turn lead to more paying members such as me.
I hope that we are able to collaborate, but if that is not possible, no hard feelings either!
Cheers,
Haris
Not a musician here, but an AI master student instead. I am currently doing my master project and I was hoping if we could help each other out.
To give you an idea what I am working on; it can be best described as a sort of dynamic low- and high-pass filter. The purpose of this system is to separate piano recordings by the playing hands. Simply put, an audio file of a piano recording comes in, two audio files come out. One contains the left hand performance, the other contains the right hand performance.
Considering that such a task (piano hand separation) has not been performed yet on raw audio, I wanted to keep it as feasible as possible by at first focusing on Boogie Woogie only. My assumption here is that Boogie Woogie would be the easiest genre to separate by playing hand due to the strong contrast between left- and right-hand performance that already exist.
The development of such systems does require data however, piano recordings in this case. I would need isolated recordings of left-hand and right-hand performance (playing Boogie Woogie). These are essential for evaluation purposes, but also for training a supervised Machine Learning approach (basically learning patterns from examples). As you might imagine however, such data does not exist. Therefore, a dataset has to be built.
This is where Wikiloops comes in!
So what would I exactly need? I am very aware that this a lot to ask, but I am aiming for at least around 50 recordings where the left- and right-hand are isolated in separate files. The reason being for why I would require so many recordings is mainly for training purposes. Supervised Machine Learning systems really do require many examples such to make a good generalization of the data it is training on. The timeline to finish this dataset can be easily spread out across 3 months however, as I can start off with a smaller dataset for prototyping and just scale up whenever the data is available.
I hope that there are member who are willing to lend out a helping hand. Not only out of sheer good-will however, because I want to propose something as well. Consider that the availability of good data is perhaps one of the greatest challenges in Machine Learning nowadays, this includes the field of music processing and blind audio source separation. The positive impact additional data has had on performance can for instance be seen in the MUSDB18 4-track music source separation task. MUSDB18 is a dataset of 150 songs where each song is separated in 4 tracks, namely: vocals, drums, bass and other. The purpose of this task therefore is to separate these tracks as good as possible. When looking at the best performing systems on this tasks (https://paperswithcode.com/sota/music-source-separation-on-musdb18), the top two were trained on additional data beyond the initial 150 tracks from the MUSDB18 dataset itself.
Now, consider that Wikiloops has roughly 151K backing tracks WITH EVERY INSTRUMENT SEPARATED IN ITS OWN TRACK, this website is in my opinion really a gold mine for AI music scientists to be discovered. If we manage to build this dataset, I will make sure to credit Wikiloops appropriately in any academic work I am going to publish.
I know that "promising exposure" is a laughed at cliche for artists nowadays, however the fact is that increased exposure of Wikiloops to the scientific community could lead to more traffic on Wikiloops and in turn lead to more paying members such as me.
I hope that we are able to collaborate, but if that is not possible, no hard feelings either!
Cheers,
Haris
+4