🥁 Human-Like Drum Rhythm Generation with LSTMs

This project explores the generation of expressive, human-like drum patterns using a sequence-to-sequence LSTM architecture trained on the Expanded Groove MIDI Dataset (E-GMD).

The pipeline is fully configuration-driven via a central YAML config file and includes:

• MIDI preprocessing and tokenization
• Training of LSTM models with varying capacities
• Token prediction and MIDI regeneration
• Evaluation via statistical metrics and visual/audio inspection

Listen to generated samples:

• Blues
• Dance
• Neworleans

Table of Contents

• Project Goals
• Architecture
  • Preprocessing
  • Model
  • Training
  • Evaluation
• Results
  • Quantitative Performance
  • Generated Samples
• Installation

Project Goals

• Model stylistic timing and velocity nuances of human drummers
• Compare model sizes (small, medium, large) and learning rates
• Evaluate generation quality both quantitatively and qualitatively
• Generate drum patterns for 12 genres:
  afrocuban, dance, blues, funk, hiphop, jazz, latin, neworleans, pop, reggae, rock, soul

Architecture

Preprocessing

1. MIDI Reading: Parse MIDI files
2. Pitch Grouping: Map MIDI pitches to 13 canonical drum categories
3. Drumroll Conversion: Convert to (T × 13) matrix
4. Segmentation: Extract overlapping sequences of segment_len
5. Tokenization: Convert each token_len-timestep beat into a token
6. Pruning: Keep top 95% most frequent tokens, rest → [UNK]

Example tokens after preprocessing is done:

Data distribution after preprocessing is done:

Model

Seq2Seq LSTM Architecture:

• Encoder: Processes input sequence with embeddings for tokens, position (cyclic sin/cos), genre, and BPM
• Decoder: Generates output tokens autoregressively with teacher forcing during training
• Embeddings: Token (64-256d), Position (4-16d), Genre (8-16d), BPM (4d)
• LSTM Layers: 1-3 layers with 128-512 hidden units depending on model size

Model	Hidden	Layers	Token Embed	Params
Small	128	1	64	~500K
Medium	256	2	128	~2M
Large	512	3	256	~8M

Training

• Loss: Cross-entropy with class weighting for imbalanced tokens
• Optimizer: AdamW with configurable learning rate
• Teacher Forcing: Linear decay from 1.0 to -2.0 over training
• Early Stopping: Patience of 5 epochs on validation loss
• Batch Size: 1024 sequences
• Epochs: Up to 500 with early stopping

All parameters configurable in config.yaml

Evaluation

Metrics:

• Token-level accuracy, precision, recall, F1-score
• Visual inspection of generated patterns
• Audio playback of MIDI outputs

Results

Quantitative Performance

Example metrics (Small model, lr=0.001, seg_512):

Metric	Value
Accuracy	0.949
F1 (macro)	0.988
Precision (macro)	0.982
Recall (macro)	0.999

Comparisons of different learningrates per model:

Comparisons of different models with learningrate of 0.001:

Generated Samples

Example of generated output:

Installation

# Clone the repository
git clone https://github.com/Arthurreuss/NN_Drumming_Project.git
cd NN_Drumming_Project

#Set up virtuall environment
python -m venv NAME
source NAME/bin/activate

# Install required Python packages
pip install -r requirements.txt

# Download the E-GMD dataset
# https://magenta.tensorflow.org/datasets/e-gmd#download
# Place it here:
# → dataset/e-gmd-v1.0.0/

# Quick start
python -m main --model small --learning_rate 0.001