🎤Vocal Percussion Classifier

This project presents a comparative study of machine learning approaches for classifying vocal percussion sounds, commonly used in beatboxing and vocal percussion performance.

Project Goals

The primary objective was to develop and evaluate different machine learning models for accurately classifying various vocal percussion sounds. We focused on creating a robust system that could distinguish between different types of vocal percussion sounds (such as kick drums, snares, hi-hats, and cymbals) with high accuracy.

Methodology & Results

We implemented and compared several machine learning approaches:

• Traditional ML Models: SVM and Random Forest classifiers using handcrafted audio features
• Deep Learning: CNN and LSTM architectures for end-to-end learning
• Hybrid Approach: Combining traditional features with neural networks

Key findings showed that:

• The CNN model achieved the highest accuracy (92%) for isolated sounds
• LSTM performed better for continuous sequences
• Traditional ML models were more computationally efficient

Technical Implementation

• Feature Extraction: Used librosa for MFCCs, spectral features, and temporal characteristics
• Data Processing: Implemented custom augmentation techniques for limited training data
• Model Training: Utilized scikit-learn for traditional ML and TensorFlow for deep learning models