Audio-Audit is a learning-focused machine learning project that analyzes short respiratory audio clips and classifies them into meaningful categories such as Cough, Heavy Breathing, Normal Breathing, and Background Noise.
This project was built as part of my journey from understanding the basics of machine learning to developing and deploying a complete, end-to-end system.
- Accepts short (4–5 second) audio recordings
- Converts audio into Mel Spectrograms
- Uses a CNN-based deep learning model for classification
- Applies confidence-based filtering to avoid uncertain predictions
- Returns "Unclear / Unknown" when confidence is low
- Provides a simple Streamlit web interface for real-world testing
- Audio preprocessing using Librosa
- Fixed-length normalization (4 seconds)
- Feature extraction via Mel Spectrograms
- CNN trained on real-world respiratory and noise samples
- Confidence thresholding to reduce false positives
This system is not a medical diagnostic tool and is intended purely for learning and experimentation.
- Sharp transient sounds (finger snaps, clicks) can sometimes resemble cough patterns
- Real-world audio is highly noisy and unpredictable
- Model performance is limited by dataset size and diversity
These observations helped me better understand the challenges of deploying ML systems outside controlled environments.
- Python
- TensorFlow / Keras
- Librosa
- NumPy
- Streamlit
- HuggingFace Spaces
- Docker
The application is deployed using HuggingFace Spaces and runs inside a Docker container.
Users can:
- Upload an audio file
- Record audio directly
- View predictions with confidence scores
This project represents my transition from learning ML concepts to:
- Working with real-world data
- Performing basic audio EDA
- Training and evaluating neural networks
- Deploying an interactive ML application
I plan to continue improving this system with more diverse data and stronger robustness.
This project is not intended for medical use.
Results should not be interpreted as health advice.
Built with curiosity, mistakes, and learning.