DSP-Based Music Genre Classifier

---

Abstract

This project presents a real-time DSP-based music genre classification system that combines handcrafted feature extraction with deep learning techniques. The system integrates classical machine learning models—K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Boosted Trees—with a lightweight Convolutional Neural Network (CNN) trained on mel-spectrograms. The hybrid model achieves high accuracy and low latency suitable for real-time applications such as music streaming platforms, DJ software, and media organization tools.

Problem Statement

Classical approaches used features like MFCCs and models such as k-NN or SVM to achieve ~70% accuracy. With the rise of deep learning, CNNs operating on spectrogram images boosted performance beyond 85%. However, real-time constraints limit deep model usage on low-resource platforms. This work proposes a hybrid DSP + CNN solution to achieve both speed and performance. So the goal is to design a music genre classification system that balances:

• High accuracy using deep learning
• Low latency using classical models
• Real-time feasibility for live applications

System Overview

DSP Features Extracted

• Mel-Frequency Cepstral Coefficients (MFCCs) – captures timbral textures
• Spectral Descriptors – centroid, rolloff, flux, ZCR
• Rhythmic Features – beat strength, BPM, tempo variance
• Harmonic Features – pitch histograms, chroma vectors

---

CNN on Mel-Spectrograms

• Preprocessing:

  • Frame audio with 2048-sample windows, 50% overlap
  • Apply FFT, Mel filterbanks, and log compression
  • Generate 64×128 mel-spectrogram image for 3s audio clips

• CNN Architecture:

  • Two Conv layers (64 filters, 3×3 and 3×5)
  • Max-pooling (2×4)
  • One dense layer (32 units) + Softmax
  • Dropout: 0.3, Optimizer: Adam (lr = 0.001)

Performance Comparison

Model	Accuracy	Precision	Recall
k-NN	60.5%	0.61	0.60
SVM	68.7%	0.70	0.69
Boosted Trees	75.4%	0.76	0.75
CNN	83.0%	0.84	0.83
Hybrid Model	85.0%	0.86	0.85

The hybrid model provides the best overall performance with low latency and high generalization.

Folder Structure

---

How to Run

Requirements

• MATLAB R2020 or newer
• Signal Processing Toolbox
• Deep Learning Toolbox.

Steps

1. Clone the repository:

git clone https://github.com/Awais-Asghar/DSP-Based-Music-Genre-Classifier

2. Open MATLAB and add all folders to path:

addpath(genpath('DSP-Based-Music-Genre-Classifier'))

3. Run any model:

run('code/KNN.m')
run('code/SVM.m')
run('code/Boosted_Trees.m')
run('CNN/cnn.m')

Dataset

• Based on Free Music Archive (FMA) dataset
• Genres used: Folk, Hip-Hop, Instrumental, International
• Audio clips are processed into 3-second segments for classification

---

Visual Results

Top influential features include:

• MFCC₂
• Spectral Flux
• Spectral Kurtosis
• Harmonic Ratio
• MFCC₃

These features play a critical role in effectively distinguishing between music genres.

Future Work

• Extend to sub-genres and multilingual music
• Add LSTM/GRU layers for temporal modeling
• Optimize CNN using quantization/pruning for embedded systems

Authors

• Awais Asghar
• Muhammad Ashar Javid
• Muhammad Hammad Sarwar
• Huzaifa Ahmad
  Department of Electrical Engineering,
  National University of Sciences and Technology (NUST), Islamabad, Pakistan

---