Sound Anomaly Detection 🎵🔍

A deep learning project for detecting anomalous sounds (e.g., footsteps, breaking glass, screams) in audio streams using transformer-based models.

📜 Project Overview

Objective

Develop a robust system to identify unusual or dangerous sound events in real-world environments using state-of-the-art audio models.

Target Anomalies:

• Footsteps in restricted areas
• Breaking glass
• Human screams/shouts
• Other context-specific anomalies

Core Components

📂 Dataset

Kaggle Dataset: AudioAnomalyDataset

• Contains labeled audio samples of:
  • Normal sounds (background noise, conversations)
  • Anomalous events (breaking glass, screams, footsteps)
• Format: .wav files at 16kHz sampling rate

🧠 Models

• Transformer-based models like Wav2Vec2, Whisper.

Installation and Run

Prerequisites

Before running the project, ensure you have the following installed:

• Python 3.10 or higher
• CUDA and cuDNN (if using GPU)
• Git
• Conda (Optional)

1. Clone the Repository

Clone the project repository to your local machine:

git clone https://github.com/AlinaShapiro/Sound_Anomaly_Detection.git
cd Sound_Anomaly_Detection

2. Set Up a Virtual Environment

conda create -n sound_anomaly_detection python=3.10
conda activate sound_anomaly_detection

3. Install Dependencies

pip install -r requirements.txt

4. Run

Inference

To run inference on a test subset of Audio-Anomaly-Dataset, use:

python inference.py --model_name wav2vec2

Applications

This project can be applied to both Security Systems and Smart Home Automation, as sound anomaly detection can expand the capabilities of traditional security systems. These systems often rely on motion sensors or magnetic contacts to detect intruders, but sound anomaly detection allows them to detect sounds like broken glass, footsteps, or quiet speech, which may indicate an intruder. This is particularly useful for large spaces where conventional sensors may not be able to detect intruders effectively.