Acoustic Vehicle Classifier — Engine Identification Using Signal Processing & ML

A machine learning web application that identifies vehicle types from engine audio signals using FFT-based feature extraction and a Support Vector Machine (SVM) classifier. Upload any .wav engine recording and the system predicts the vehicle class with high accuracy.

---

Live Demo

Run locally — see setup instructions below.

---

Project Overview

This project classifies vehicle engine sounds into 9 distinct vehicle categories:

Class	Description
CAR	Standard passenger car
BIKE	Motorcycle
BUS	Full-size bus
MINIBUS	Minibus / shuttle
PICKUP	Pickup truck
CROSSOVER	Crossover SUV
JEEP	Off-road vehicle
SPORTS_CAR	High-performance car
TRUCK	Heavy commercial truck

---

How It Works

Audio (.wav)
    ↓
librosa.load()         — load raw signal
    ↓
FFT (numpy)            — convert time → frequency domain
    ↓
Feature Extraction     — 21 features extracted
    ↓
StandardScaler         — normalize features
    ↓
SVM Classifier         — RBF kernel, C=10
    ↓
Vehicle Prediction     — with certainty %

---

Features Extracted (21 total)

#	Feature	Description
1	Spectral Centroid	Center of mass of frequency spectrum
2	Total Energy	Overall signal power
3	Dominant Frequency	Peak frequency in spectrum
4	Band Energy Ratio	Low vs high frequency energy ratio
5	Spectral Rolloff	Frequency below which 85% of energy exists
6	Spectral Bandwidth	Spread of the frequency spectrum
7	Zero Crossing Rate	Signal roughness indicator
8	RMS Energy	Root mean square amplitude
9–21	MFCC 1–13	Mel-frequency cepstral coefficients — engine timbre fingerprint

---

Model Performance

Accuracy     : 98.4%
CV Score     : 97.9% ± 0.4%

              precision  recall  f1-score
BIKE            1.00      1.00      1.00
BUS             1.00      1.00      1.00
CAR             0.95      0.97      0.96
CROSSOVER       0.96      0.97      0.97
JEEP            0.98      0.97      0.98
MINIBUS         1.00      0.97      0.99
PICKUP          0.98      0.96      0.97
SPORTS_CAR      1.00      1.00      1.00
TRUCK           1.00      1.00      1.00

---

Tech Stack

Layer	Technology
Backend	Python, Flask
Signal Processing	NumPy (FFT), Librosa
Machine Learning	Scikit-learn (SVM, StandardScaler)
Visualization	Matplotlib
Frontend	HTML, CSS, Vanilla JS
Fonts	Bebas Neue, Share Tech Mono

---

Project Evolution

Phase	What Changed
Phase 1	Started with synthetic generated audio — 2 classes (Car vs Bike) — 4 features — ~70% accuracy
Phase 2	Switched to real VISC8 dataset — expanded to 9 vehicle classes — 4 features - ~81% accuracy
Phase 3	Upgraded to 21 features including MFCCs — accuracy jumped to 98.4%

---

Dataset

This project uses the VISC8 Dataset — a benchmark dataset of 5,980 vehicle interior engine sounds recorded at 48kHz.

Download: https://zenodo.org/records/5606504

After downloading, extract and organize the files like this:

vehicle_classifier/
└── dataset/
    ├── car/
    ├── bike/
    ├── bus/
    ├── minibus/
    ├── pickup/
    ├── crossover/
    ├── jeep/
    ├── sports_car/
    └── truck/

---

Installation & Setup

1. Clone the repository

git clone https://github.com/thavasix-gr8/engine-identification-acoustic-ml.git
cd engine-identification-acoustic-ml/vehicle_classifier

2. Install dependencies

pip install flask librosa numpy scikit-learn matplotlib soundfile

3. Download and prepare the dataset

Download VISC8 from zenodo.org/records/5606504 and place audio files into the folder structure shown above.

4. Train the model

python train_model.py

This generates model.pkl and scaler.pkl in the project root.

5. Run the web app

python app.py

Open your browser at http://127.0.0.1:5000

---

Usage

1. Open the web app in your browser
2. Drag and drop any .wav engine audio file into the upload zone
3. The system will display:
   • Vehicle type prediction with certainty percentage
   • 4 key spectral feature values
   • Time domain waveform graph
   • Frequency domain FFT spectrum graph
4. Click ANALYZE_NEW_FILE to test another file

---

File Structure

vehicle_classifier/
├── dataset/              ← audio files (not included, download separately)
├── static/
│   ├── main.js           ← frontend logic
│   └── style.css         ← styling
├── templates/
│   └── index.html        ← web interface
├── app.py                ← Flask backend + feature extraction
├── extract_features.py   ← feature extraction module
├── train_model.py        ← model training script
├── predict.py            ← terminal prediction script
└── generate_data.py      ← synthetic data generator (for testing)

---

License

MIT License — free to use for academic and research purposes.

---

Built as a Signal Processing & Machine Learning mini project.