🎧 Multiclass Audio Event Detection with Deep Learning

A robust pipeline for multiclass audio event detection using deep learning. The system classifies audio into four classes:

• 🔫 Gunshot
• 🚗 Vehicle
• 🪓 Wood_Cutting
• ✅ NonThreat

It includes real-time alerting logic (e.g., via LoRa) and deployment-ready TensorFlow Lite models for edge devices.

📁 Dataset Structure

org_dataset/
├── Gunshot/
│   ├── file1.wav
│   └── ...
├── Vehicle/
│   ├── file1.wav
│   └── ...
├── Wood_Cutting/
│   ├── file1.wav
│   └── ...
└── NonThreat/
    ├── file1.wav
    └── ...

• Format: .wav files per class
• Holdout Set: A separate folder (e.g., realtest/) with test samples from each class

🚀 Features

• 📊 EDA: Class distribution visualization and spectrogram previews
• 🧹 Preprocessing: Converts audio to log-mel spectrograms; ready for CNN input
• 🧠 Model:
  • Data augmentation
  • Batch normalization
  • Dropout
  • L2 regularization
• 🏋️ Training:
  • Early stopping
  • Learning rate reduction
  • Checkpointing best model
• ✅ Evaluation: Confusion matrix & classification report
• 🛠️ Deployment: Converted to TFLite for edge inference

🛠️ Installation

1. Clone the Repository

git clone https://github.com/yourusername/audio-event-detection.git
cd audio-event-detection

2. Create Virtual Environment & Install Dependencies

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

📦 Main Dependencies

• tensorflow
• librosa
• numpy
• matplotlib
• scikit-learn

📂 Dataset Setup

• Place your dataset in the structure mentioned above
• Update dataset paths in the scripts if needed

🔧 Customization

• Class Names: Update the class_names list in scripts for custom classes
• Alert Logic: Modify send_lora_alert() in tflite_inference_lora.py
• Paths: Update dataset/model paths as per your directory setup

🧠 Model Architecture

ConvBlock x3 (Conv2D → BatchNorm → MaxPooling)
↓
Data Augmentation (random translation, zoom)
↓
GlobalAveragePooling2D
↓
Dense Layer → Dropout
↓
Output Layer (Softmax)

📸 Refer to cnn_model_architecture.png for a diagrammatic view.

📊 Results

• Validation Accuracy: ~96% on holdout set
• Evaluation Metrics: Confusion Matrix, Classification Report (from test_holdout.py)

📱 Edge Deployment

• Model File: multiclass_model.tflite (~122 KB)
• Deploy to: Raspberry Pi, microcontrollers, IoT devices
• Real-Time Detection: Included example with alerting logic

🙌 Credits

Built with 💡 using TensorFlow, Librosa, and open-source tools.