125 points by musicianai 1 year ago flag hide 16 comments
johnsmith 1 year ago next
Interesting case study! I've been playing around with generative music algorithms recently. I find it fascinating how RNNs can learn patterns in data and generate new content based on that learning. Well written and easy to understand.
codeninja 1 year ago next
@johnsmith, have you looked into any specific music composition libraries or frameworks? I'm looking to get started with generative music and was wondering what your recommendations would be.
johnsmith 1 year ago prev next
@codeninja, I've been using the Sonic Pi framework for my experiments. It's pretty easy to use and has a lot of documentation available. I think it would be a great starting point for generating music with RNNs.
codeninja 1 year ago next
@johnsmith, thanks for the recommendation. I'll definitely check out Sonic Pi and see if it fits my needs.
anotheruser 1 year ago prev next
Great write-up, really enjoyed reading this. I wonder how these generative music algorithms would compare to machine-learning generated art and other creative fields?
musicgenius 1 year ago prev next
I've been working on a generative music app that uses RNNs to generate accompaniment for user-entered melodies. It's been a fascinating exploration of RNN architectures and optimization techniques.
anotheruser 1 year ago next
@musicgenius, that's really cool. How do you handle user input and RNN output synchronization?
musicgenius 1 year ago next
@anotheruser, I use a small real-time audio processing library to handle user input and RNN output synchronization. It's not perfect, but it gets the job done.
aiengineer 1 year ago prev next
Generative music is a fascinating application of RNNs. I'm working on a project that uses a reinforcement learning approach to train a music composition RNN.
deepthinker 1 year ago next
@aiengineer, that's really interesting. How do you approach reward design for your reinforcement learning system?
aiengineer 1 year ago next
@deepthinker, I use a combination of pre-defined musical rules and learned preferences to design rewards. It's an iterative process and it required a lot of fine-tuning to get it right.
johnsmith 1 year ago prev next
@aiengineer, that sounds really cool. I'd love to learn more about your RL system and how it compares to the more traditional supervised learning approach.
machinelearner 1 year ago prev next
I've been working on a generative poetry app that uses a similar reinforcement learning approach. I find it fascinating to see how similar techniques can be applied to such different domains.
codeninja 1 year ago next
@machinelearner, that's amazing. Can you share more details about the architecture and optimization techniques you used in your generative poetry app?
computationfive 1 year ago prev next
In my music generation experiments, I've found that adding noise to the inputs can give the model more unpredictability and complexity.
musicgenius 1 year ago next
@computationfive, I've been trying something similar with my generative music app. I've found that adding noise not only adds unpredictability, but can also help avoid repetition of musical patterns.