Music Generation with RNNs
Deep learning system generating Irish folk music using 2-Layer LSTM trained on 214K tunes
Overview
Implemented a music generation system using Recurrent Neural Networks (RNNs) trained on the IrishMAN dataset containing 214,122 Irish folk tunes in ABC notation format. The model learns to generate new music by predicting the next character in a sequence, similar to language models but applied to music. Achieved 74.91% top-1 accuracy and 95.40% top-5 accuracy. Built a professional Flask web application with gradient-based UI, real-time generation, ABC notation display, and generation history tracking. The system successfully generates syntactically valid ABC notation tunes that follow musical conventions learned from the training data.
Problem
Music generation requires understanding complex sequential patterns and musical conventions. Traditional approaches struggle with variable-length sequences and maintaining musical coherence.
Solution
Implemented 2-Layer LSTM with 256 hidden units and 128-dimensional embeddings, trained on 214K Irish folk tunes. Applied gradient clipping for stability, dropout for regularization, and learning rate scheduling. Built Flask web application with real-time generation, temperature control, and ABC notation validation.
Results
Top-1 Accuracy: 74.91%
Top-5 Accuracy: 95.40%
Model Parameters: 959,715
Training Loss: 0.7918
Validation Loss: 0.7729
Trained on 214,122 Irish folk tunes
Generates syntactically valid ABC notation
Professional web interface with real-time generation
Temperature control for creativity adjustment
Generation history tracking and ABC download