Dynamic Music Composition System Using LSTM and Transformer Networks

This paper introduces a music composition system using AI democratizing creative expression by leveraging deep learning technologies. By utilizing LSTM and Transformer neural networks trained on MIDI datasets, our system produces unique compositions from user input. The architecture employs a Streamlit frontend, Flask/FastAPI backend, and sophisticated music processing pipeline. Users describe mood, genre, tempo, and instrumentation using an easy-to-use interface, with real-time-generated compositions. Feature extraction operations transform MIDI into learnable representations, and Qdrant vector storage facilitates similarity-based inspiration. Real-time adaptation improves compositions recursively based on user feedback. Experimental benchmarking with 85 subjects showed 87% satisfaction with mood representation and 92% accuracy with genre classification, with average generation times of 3.8 seconds for 30-second compositions. This work promotes computational creativity through the integration of advanced AI models with user-friendly interfaces, changing the way non-musicians interact with music composition without sacrificing theoretical integrity.