🎵 Audio Classifier

A binary classification model that labels 10-second audio segments from full songs as either "good" (1) or "bad" (0) for use in audio-based applications like guessing games, highlight extraction, or clip curation.

---

🔍 Overview

This project powers automated clip selection for Bollyguess, improving the quality of daily audio snippets by predicting which segments are likely to be familiar yet challenging.

The model is trained on manually labeled segments and extracts audio features such as MFCC, chroma, tempo, and pitch using Librosa. It uses a feedforward neural network built in PyTorch, and is evaluated with standard classification metrics.

---

🛠 Tech Stack

• Python
• PyTorch – model architecture and training
• TensorFlow – experimentation support
• Librosa – feature extraction (MFCC, chroma, tempo, pitch)
• FFmpeg – audio slicing and preprocessing
• Scikit-learn – evaluation (confusion matrix, F1-score)
• Pandas

---

🧠 Model Architecture

• 4 hidden layers with ReLU activation
• Final layer uses Sigmoid for binary classification
• Trained on labeled 10-second audio clips (0 = not suitable, 1 = suitable)

---

⚙️ Pipeline

1. Audio Preprocessing
   Full songs are sliced into 10-second segments using FFmpeg.

2. Feature Extraction
   Each segment is converted to a feature vector using Librosa:

   • MFCCs
   • Chroma
   • Tempo
   • Pitch

3. Model Training
   A binary classifier is trained on the extracted features using PyTorch.

4. Evaluation
   Model performance is analyzed using:

   • F1-score
   • Confusion matrix

5. Deployment (Optional)
   The classifier can be integrated into apps for automated segment selection.

---

📊 Example Use Case

Used in Bollyguess to select ideal audio clips for a daily Bollywood music guessing game.

---

🚀 Getting Started

1. Clone the repo

   git clone https://github.com/omn25/audio-classifier.git
   cd audio-classifier

2. Install dependencies

   pip install -r requirements.txt

3. Run preprocessing

   python model_trainer/build_dataset.py --input songs/ --output clips/

4. Extract features

   python model_trainer/utils.py --input clips/ --output features.csv

5. Train the model

   python model_trainer/train.py --features features.csv

---

🧪 Sample Output

• Segment: 00:50–01:00 → Predicted: 1
• Segment: 02:15–02:25 → Predicted: 0

---

📂 Project Structure

audio-classifier/
│
├── model_trainer/
│   ├── build_dataset.py
│   ├── model.py
│   ├── test_adaptive_segments.py
│   ├── test_model.py
│   ├── train.py
│   └── utils.py
│
├── songs/                  # Optional: raw input songs (if used)
├── test_songs/            # Sample songs for testing predictions
│   ├── abhi_na_jao_chod_kar.mp3
│   └── ghar_more_pardesiya.mp3

---

📄 License

MIT License

---

🙋‍♂️ Contact

Built by Om Nathwani
Email: ornathwa@uwaterloo.ca
GitHub: omn25