🔊 Urban Sound Classifier - Deep Learning with PyTorch

This project implements a full pipeline for preprocessing, modeling, and metric visualization for urban sound classification using PyTorch and the UrbanSound8K dataset. The pipeline handles audio files, applies data augmentation techniques, and converts data into MEL spectrograms ready to feed into a CNN.

---

📁 Project Structure

.
├── checkpoint/                     # Checkpoints with models, metrics, scheduler and optimizer
│   └── train_and_val_metrics.png   # Plot of accuracy and loss
├── data analysis/                  # Notebook or scripts with data preprocessing analysis
├── src/                            # Source code
│   ├── inference.py                # Inference routine for trained models
│   ├── model.py                    # CNN architecture
│   ├── training.py                 # Training loop, validation, early stopping
│   └── utils.py                    # Dataset class, preprocessing functions
├── UrbanSound8K/                   # Dataset folder
├── ForPrediction.py                # Script for inference on new audio files
├── UrbanSound_Training.py          # Main training routine
└── README.md

---

📚 Objective

The goal of this project is to develop a classifier for urban sounds using convolutional neural networks (CNNs) with PyTorch. Sounds are extracted from the UrbanSound8K dataset, and the system can identify noise such as car horns, dog barks, sirens, and more, based on MEL spectrograms. The pipeline is complete: from raw audio loading to CNN training and results visualization.

---

🔄 Preprocessing Pipeline

• 🎵 Reading .wav files using torchaudio
• 🔁 Transformation into MelSpectrogram with configurable parameters
• 🧪 Application of SpecAugment (time and frequency masking)
• 🔢 Normalization of spectrogram data
• 🏷️ Conversion to tensors and label pairing

Each sample is standardized to a fixed input size, making CNN training consistent.

---

📦 Custom Dataset

Custom dataset based on torch.utils.data.Dataset and UrbanSound8K:

• Reads from metadata/UrbanSound8K.csv
• Uses the fold column to split training and validation sets
• Lazy loading of .wav files
• Spectrogram normalization and caching
• Conditional data augmentation only during training

---

🏋️ Training

Model training is handled by the Trainer class, which includes:

• ✅ Support for early stopping and automatic checkpoints
• 📉 Calculation of metrics like loss, accuracy, recall, and F1
• 📝 Logs saved in .json formats
• 📊 Automatic plotting of training curves (loss and accuracy)
• 🧪 Validation at the end of each epoch

CNN architecture includes:

• 🔹 4 convolutional blocks with BatchNorm, ReLU, Dropout
• 🔹 MaxPooling between blocks
• 🔹 Flatten + fully connected layers
• 🔹 Final Softmax layer for 10-class classification

---

📊 Metrics Visualization

Automatic visualizations after training:

• Metric logs saved per epoch as CSV files
• Visualization script in utils/visualization.py
• Charts for:
  • 🎯 Accuracy and loss per epoch
  • 🔄 Execution time per epoch

---

🎼 Dataset

Using the UrbanSound8K dataset:

• 🔊 8732 audio files (.wav)
• 🏷️ 10 classes of urban sounds (e.g., siren, bark, car horn)
• 📁 Split into 10 folders (fold1 to fold10)
• 🗂️ Metadata file metadata/UrbanSound8K.csv contains:
  • slice_file_name
  • fold
  • classID

🔗 Download link: UrbanSound8K

---

📁 Requirements

Install the requirements with:

pip install -r requirements.txt

Key libraries:

• torch
• torchaudio
• scikit-learn
• matplotlib
• tqdm
• pandas
• numpy

---

📌 References

• UrbanSound8K Dataset
• SpecAugment: Data Augmentation for ASR
• PyTorch
• Torchaudio Docs
• Scikit-learn Metrics
• Audio Deep Learning Made Simple - Ketan Doshi

---

👨‍💻 Author

Developed by Lucas Alves
📧 Email: alves_lucasoliveira@usp.br
🐙 GitHub: cyblx
💼 LinkedIn: cyblx