70 points by music_creator 5 months ago flag hide 12 comments
deeplearningmusic 5 months ago next
Excited to share our latest project - Handwritten Music Composition using AI and Deep Learning! We've trained a model to generate musical notation and it's been a fascinating journey learning about the intricacies of musical theory and how AI can augment the creative process.
ml4art 5 months ago next
Wow, I can't wait to try it out! Do you have any plans to open-source the model or make it accessible as a web app?
openthinking 5 months ago next
@ml4art You're in luck! Our web app is launching next week. Stay tuned for updates and a link to the public beta!
musictheorynerd 5 months ago prev next
@deeplearningmusic This is so cool! I've been following your work and it's great to see the progress you've made. How did you handle symbol recognition and ensuring the compositions make sense musically?
deeplearningmusic 5 months ago next
@musictheorynerd Thank you! We used semantic segmentation to identify and classify symbols, and incorporated music theory rules to ensure the generated compositions are playable. We'll definitely consider open-sourcing or creating a web app for greater accessibility.
aiartlover 5 months ago next
@openthinking That's amazing! I'm looking forward to testing it out. How do you plan to manage copyright and ownership since the compositions are generated by AI?
openthinking 5 months ago next
@aiartlover Since the compositions are generated by the AI, there is no human authorship, and the output is not protected by copyright. We encourage users to build upon the generated compositions and to create their own amazing musical ideas!
algorithmwiz 5 months ago prev next
@deeplearningmusic That's a great combination of techniques. How did you ensure thegenerated compositions don't just follow the rules, but actually sound pleasant to listen to?
deeplearningmusic 5 months ago next
@algorithmwiz We modeled our approach on how human composers learn and incorporate music theory. After generating a composition, we had human musicians evaluate and adjust the compositions to make them more enjoyable.
curioushomemusician 5 months ago prev next
I'm curious, how big was your dataset and how did you approach training the model?
deeplearningmusic 5 months ago next
@curioushomemusician Our dataset consisted of over 10,000 pieces of sheet music and we used a combination of LSTM and Transformer architectures to train the model. This enabled it to capture long-term dependencies in the musical notation while also maintaining the broader context of the piece.
ml4musician 5 months ago prev next
@curioushomemusician We also used techniques like curriculum learning and minibatch shuffling to stabilize training, improve convergence, and generate more diverse compositions.