Environmental Sound Classification Using Machine Learning

This project was developed for the GreenWave TechLabs Machine Learning Internship technical screening task.

Objective

The objective is to classify environmental audio into four categories:

• Car Horn
• Dog Bark
• Engine Sound
• Rain

Dataset

The project uses selected classes from the ESC-50 Environmental Sound Dataset.

The sort_dataset.py script reads the ESC-50 metadata and organises the selected audio files into class folders.

Feature Extraction

The project uses 13 Mel-Frequency Cepstral Coefficients (MFCCs).

MFCCs were selected because they provide a compact representation of the frequency characteristics of audio and approximate how humans perceive sound.

Model

A Random Forest Classifier with 200 decision trees was used.

Random Forest was selected because:

• It works well with MFCC features.
• It trains quickly on a standard laptop.
• It does not require a GPU.
• It provides feature importance.
• It is easy to explain and demonstrate.

Model Evaluation

The final Random Forest model was evaluated on four environmental sound classes.

Metric	Score
Accuracy	84.38%
Precision	87.78%
Recall	84.38%
F1-score	85.06%

Class-wise Performance

Sound Class	Precision	Recall	F1-score
Car Horn	88%	88%	88%
Dog Bark	100%	88%	93%
Engine Sound	64%	88%	74%
Rain	88%	62%	73%

The model was evaluated using accuracy, precision, recall, F1-score, and a confusion matrix.

Confusion Matrix

Feature Importance

Bonus Features

• Data augmentation
• Noise robustness testing
• Real-time microphone prediction
• Streamlit web interface
• Confidence score display

Installation

pip install -r requirements.txt

## How to Run

### Launch the Streamlit Web App

```bash
streamlit run app.py

Predict a Saved Audio File

python predict_single_file.py

Use Live Microphone Prediction

python predict_live_mic.py

Retrain the Model

python audio_classifier.py