65 points by music_ai_researcher 7 months ago flag hide 13 comments
johnsmith 7 months ago next
Fascinating topic! I wonder how far off we are from AI created music being as good as human composers.
musictech 7 months ago next
@johnsmith I believe that day will come, but for now, there's still a certain human touch that can't be fully replicated.
ai_enthusiast 7 months ago prev next
I've been working on a similar project in my spare time. It seems promising, but there are still many challenges to overcome.
musictech 7 months ago next
@ai_enthusiast That's really interesting. What challenges have you faced so far?
johnsmith 7 months ago prev next
@ai_enthusiast Are you sharing your findings publicly? I'd love to see what you've done so far.
ai_enthusiast 7 months ago prev next
@musictech The complexity of music is the main challenge. There are various elements to consider, such as melody, harmony, rhythm, and structure. Getting all these components to work seamlessly together is the most challenging part. @johnsmith I haven't yet, but I plan to once I make more progress.
codewiz 7 months ago next
@ai_enthusiast That's true, there's a lot to consider. Have you tried using any available datasets for training models? I've heard there are some sizeable datasets specifically for music composition.
ai_enthusiast 7 months ago next
@codewiz Yes, I've tried a few datasets. A difficulty with them is curation. Many datasets contain elements that don't necessarily work well together, making the model's output less musically appealing. It may be better to create a curated dataset specifically for music generation if resources allow.
quantstart 7 months ago prev next
@johnsmith I've also been working on a similar project. Getting the melody right while ensuring the harmonies work well together has proven difficult.
johnsmith 7 months ago next
@quantstart It's always great to meet others interested in this topic. Have you attempted combining LSTM and GRU cells to tackle the melody and harmony aspects of your project? I found it had promising results in my early tests.
quantstart 7 months ago prev next
@johnsmith I haven't tried that specific combination. I will do some more research on this and see how it can improve my results. Thanks for the tip!
ml_prodigy 7 months ago prev next
Has anyone tried using reinforcement learning instead of supervised methods for musical AI composition? The ability to learn from multiple iterations while self-correcting might yield more pleasing results.
codewiz 7 months ago next
@ml_prodigy I've experimented with RL techniques, but with limited success so far. RL methods need more compute resources compared to supervised techniques, and finding the optimal balance between exploration and exploitation remains a challenge. I will keep trying though!