🚨 Simple Audio Classification Using Feature Extraction

EC5011 – Task 02

This project focuses on classifying audio signals using digital signal processing (DSP) techniques. It involves two major tasks: Part 1: Classifying unknown audio samples into class_1 or class_2 using MFCC and FFT features. Part 2: Differentiating ambulance sirens and firetruck sirens using FFT-based spectral analysis, Butterworth bandpass filtering, and energy-based features.

📘 Introduction

Audio classification enables machines to identify and distinguish between sounds in real time. This project uses feature extraction techniques such as MFCC, FFT, and energy-based filtering to classify emergency sirens. The project has two main objectives: 🔹 Part 1: Basic Audio Classification

Classify unknown audio samples into class_1 or class_2 Use MFCC and FFT features Compare samples using: Euclidean Distance Manhattan Distance Cosine Similarity

Predict class based on best similarity score 🔹 Part 2: Ambulance vs Fire Truck Siren Classification Perform spectral analysis to find distinguishing frequency bands Design 4th-order Butterworth bandpass filters Compute log-energy features from each band Combine MFCC + Energy features

Use ensemble classifier (bagging) to classify siren type

Achieved ≈83.08% accuracy

🔧 Tools & Technologies Used MATLAB R2023 / Signal Processing Toolbox MFCC extraction (mfcc()) FFT spectrum analysis

Butterworth filter design (butter()) Machine learning ensemble classifier (fitensemble())

📐 Methodology

1. Feature Extraction Techniques ✔ MFCC

Extract first 13 MFCC coefficients

Use mean values to form a fixed-size feature vector Effective for tonal and perceptual analysis (speech-like signals)

✔ FFT

Converts time-domain signal into frequency domain Helpful for identifying dominant spectral peaks Used for frequency-band selection in Part 2

✔ Bandpass Filtering

Three Butterworth filters were used:

Band Frequency Range (Hz) Purpose Band 1 500–1500 Captures lower components, more common in ambulance sirens Band 2 1500–3000 Transitional band Band 3 3000–4500 Higher harmonics, often stronger in firetruck sirens

✔ Energy Features

Compute energy in each filtered band Apply log transformation to stabilize dynamic ranges Combine with MFCC features for classification

🔍 Distance & Similarity Metrics (Part 1) Metric Use Euclidean Distance Measures magnitude similarity Manhattan Distance Robust to outliers Cosine Similarity Measures shape similarity, effective for audio

Final prediction is based on the class with highest similarity (or lowest distance).

🎛 Filter Design (Part 2) Filter Type:

4th-Order Butterworth Bandpass Filters

Why Butterworth? Maximally flat passband

No ripples Stable frequency response

Good for clean energy extraction

Tool Used:

MATLAB butter() function

🧠 Classification Approach ✔ Part 1

Extract MFCC or FFT features Compare feature vectors using similarity metrics Assign class based on highest similarity score

✔ Part 2

Extract MFCC + Log-Energy features Use ensemble tree classifier (bagging) Automatically learns thresholds Tested on siren audio samples

Achieved 83.08% accuracy