🌊 Marine Sound Classifier

A deep learning-based marine sound classification system built using CNN and Streamlit.
This project classifies underwater audio into whale, ship, dolphin, or unknown using spectrogram-based learning. Designed to handle real-world noisy ocean environments with robust unknown detection.

---

🚀 Features

• 🎧 Upload or record audio directly in UI
• 🧠 CNN-based classification using Mel Spectrograms
• ⚖️ Confidence-based Unknown detection
• 🔍 Chunk-wise prediction for better accuracy
• 📊 Probability breakdown for each class

---

🧠 Model Details

• Architecture: Custom CNN with BatchNorm, Dropout & Global Average Pooling
• Input: Log Mel Spectrograms (128 × 60)
• Classes:
  • Whale 🐋
  • Ship 🚢
  • Dolphin 🐬
  • Unknown ❓
• Loss: Weighted CrossEntropy (reduced bias toward unknown)

---

⚙️ Tech Stack

• Python 🐍
• PyTorch 🔥
• Librosa 🎶
• Streamlit 🌐
• NumPy

---

📁 Project StructureMarine_Sound_Classifier/

│ ├── app.py ├── marine_model.pth ├── requirements.txt ├── demo.png │ ├── models/ ├── preprocessing/ ├── data/

---

📈 Results

• High accuracy on ships and dolphins
• Improved whale detection after preprocessing tuning
• Stable unknown classification using confidence-based logic

---

🧪 How to Run

pip install -r requirements.txt
streamlit run app.py