Music Generation using Deep Learning

Project Overview

This deep learning project leverages generative sequence modeling to compose classical music. By training on a dataset of piano compositions, the recurrent neural network learns notes, chords, and temporal intervals to compose original music files saved in standard MIDI format.

Key Features & Objectives

• Generative Compositions: Composes chord sequences and melodic notes from scratch.
• Sequential Note Vectorization: Transcodes raw MIDI structures into categorical indexes.
• Adjustable Composition Variation: Controls sequence creativity using temperature sampling factors.

Technical Stack & Technologies

• Language: Python
• Deep Learning: TensorFlow, Keras
• MIDI Parsing: Music21 (musical notation toolkit)
• Sequence Formatting: NumPy, MIDI players

Core Techniques & Generative Models

• Stacked LSTM Architectures: Designed recurrent networks with stacked LSTM and Dropout layers, predicting the next note/chord index based on a trailing sequence window.
• Musical Tokenization: Parsed notes (pitches) and concurrent chords into indexed categories, managing high-dimensional softmax outputs.

• Temperature-Guided Sampling: Configured output generator functions with a temperature-scaling factor. Higher temperatures encourage creative variations and note choices, while lower values restrict choices to highly probable training notes to prevent chaotic compositions.