BEATBOX - Drum Sound Classification System

A machine learning system for classifying drum sounds using a hybrid KNN approach with personalization capabilities.

Model Comparison

Approach	Description	Accuracy / ARI	Strengths	Limitations
Unsupervised	Clustering with K-Means, GMM, HDBSCAN, Hierarchical	ARI ≈ 0.11	No labels needed, flexible discovery	Poor label alignment, low structure clarity
Seeded K-Means	K-Means initialized with one participant’s true centroids	ARI ≈ 0.13	Adds structure, improves cluster quality	Still low accuracy, depends on selected participant
Supervised	Global KNN model trained on all users, tested on unseen participants	57.0% Accuracy	Simple, interpretable	Weak generalization to new users
Hybrid	KNN with 5 user-specific samples + feature weighting via Ridge/Lasso	83.8% Accuracy	Personalized, low data requirement	Requires minimal user input at runtime

Dataset

This project uses the combined AVP-LVT Vocal Percussion Dataset, which merges:

• AVP (Amateur Vocal Percussion) — 28 performers
• LVT (Live Vocalised Transcription) — 20 performers

Both datasets are preprocessed and included in the repository under the data/ directory. These include segmented audio clips, standardized labels, and annotation metadata for training, evaluation, and synthesis.

Setup Instructions

1. Create a Python virtual environment:

python -m venv venv
source venv/bin/activate  # On macOS/Linux
# or
.\venv\Scripts\activate  # On Windows

2. Install required packages:

pip install -r requirements.txt

Workflow

1. Data Collection and Processing

   • Run datacollection.ipynb to create a master dataframe of all audio samples
   • Run audiosegmentation.ipynb to segment audio into individual drum sounds
   • Run segmentaugmentation.ipynb to create augmented segments for training

2. Feature Extraction

   • Navigate to the feature_extraction folder
   • Run all three feature extraction notebooks:
     • mfcc_extraction.ipynb
     • envelope_features.ipynb
     • feature_selection.ipynb

3. Model Training

   • Navigate to the models folder
   • Run the model files

4. Final Pipeline

   • Navigate to final_pipeline/code/
   • Run pipeline.ipynb to execute the complete system

Key Features

• Hybrid KNN model with personalization capabilities
• Feature extraction using MFCCs and envelope features
• Data augmentation for improved model robustness
• Ridge and Lasso regression for feature selection
• Visualization tools for model analysis

Model Architecture

The system uses a hybrid KNN approach that:

• Combines base training data with user-specific examples
• Uses Ridge/Lasso regression to identify important features
• Implements weighted distance metrics for improved classification
• Adapts to individual playing styles