87 points by music_lover 1 year ago flag hide 16 comments
johnsmith 1 year ago next
This is amazing! I've been looking for something like this for so long. Great work!
codeuser 1 year ago next
I know right! I've been trying to build something similar, but never had any success. What kind of data did you use for training?
johnsmith 1 year ago next
I used the the 'MusicNet' dataset, which has over 300 hours of recordings. It was a bit difficult to preprocess, but totally worth it.
mlgirl 1 year ago prev next
Wow, 300 hours of recordings sounds like a lot. Are you planning to make your preprocessing scripts available?
johnsmith 1 year ago next
Definitely! I might even include a tutorial on how to use it. I'll let you know once it's ready.
aiengineer 1 year ago prev next
This is a really cool project. I've been working on a similar problem for video analysis, and I found that the key is to pay attention to the tempo of the piece. Did you take that into account?
johnsmith 1 year ago next
Absolutely! In MusicNet, tempo is considered one of the most important features, so I made sure to focus on it. I used a simple statistical method to extract the tempo, but it worked pretty well.
quantumcoder 1 year ago prev next
One question, did you consider any data augmentation techniques while training, like time stretching or pitch shifting?
johnsmith 1 year ago next
Yes, I did. I applied both time stretching and pitch shifting with various factors to increase the variability of the pieces. It was surprisingly effective.
johnsmith 1 year ago next
Pitch shifting can be a bit tricky, but if you do it carefully, it's very effective. I used a simple method that uses a Fast Fourier Transform (FFT) to shift the pitches.
quantumcoder 1 year ago next
Sounds like an interesting method. I'll give it a try. Thanks for sharing.
codeuser 1 year ago prev next
I also found time stretching to be super helpful. I'm using it for my project as well. Have you tried pitch shifting? I've heard it can be a bit difficult to deal with.
tensorflowguru 1 year ago prev next
Nice! I've been trying to solve this problem myself for a while now. Do you mind if I take a look at your code and try to learn from it?
johnsmith 1 year ago next
Sure, feel free to take a look at the code. It's still a work in progress, but I'm happy to share it. Let me know if you have any questions.
tensorflowguru 1 year ago prev next
Thanks, I'll definitely check it out. Are you planning to release it as open source?
johnsmith 1 year ago next
Yes, definitely. I want to make it accessible to as many people as possible. I'll let you know once it's ready.