Music Generation using LSTMs

A deep learning project exploring sequence modeling for symbolic music generation using Long Short-Term Memory (LSTM) recurrent networks. The model learns pitch, timing, and temporal patterns from classical MIDI piano performances in the MAESTRO dataset to generate new musical sequences.

---

Overview

Generating realistic music requires capturing long-term dependencies, rhythmic cadence, and structural harmony. This project approaches music generation as a sequential auto-regressive task:

• Input Representation: Extracts pitch, note duration, and step time from MIDI files.
• Architecture: Multi-layer LSTM network trained to predict the next note/event given a preceding sequence context.
• Sampling & Generation: Generates new MIDI files starting from a seed prompt or random initialization.

---

Repository Structure

music-rnn/
├── data/
│   └── maestro-v1.0.0-midi/
│       └── maestro-v1.0.0/
│           ├── 2004/
│           ├── 2006/
│           ├── 2008/
│           ├── 2009/
│           ├── 2011/
│           └── ...
└── generate.py
└── music_lstm.pt

---

Dataset

This project uses the MAESTRO Dataset (v1.0.0), a collection of over 200 hours of virtuosic piano performances captured with fine pitch and timing details:

1. Download the MAESTRO v1.0.0 MIDI dataset.
2. Extract the contents into data/ as shown in the directory structure above.

---

Getting Started

Prerequisites

• Python 3.8+
• Recommended dependencies:
• torch (or tensorflow)
• pretty_midi or mido (for MIDI parsing)
• numpy
• matplotlib

Install dependencies:

pip install torch pretty_midi numpy

---

Usage

1. Generating Music

Run generate.py to sample a new sequence from a trained checkpoint and export it as a standard .mid file:

python generate.py --output output.mid --length 500 --temperature 1.0

Common Arguments

• --output: Output file path for the generated MIDI file.
• --length: Number of note events to generate.
• --temperature: Sampling temperature (higher = more experimental, lower = safer/more repetitive).

---

How It Works

1. MIDI Parsing: MIDI events are parsed into a numerical matrix where each event represents pitch, step time (delay between notes), and note duration.
2. Sequential Windowing: Training data is split into sliding sequence windows (e.g., 100 timesteps) mapped to the immediate next event.
3. Training: An LSTM processes these sequences, optimizing cross-entropy loss for pitch prediction and MSE loss for timing attributes.
4. Sampling Loop: The model predicts step $N+1$, appends it to the input window, and slides forward to generate the entire sequence iteratively.

---