🎧 Source Separation with NMF

A simple and clean implementation of audio source separation using Non-negative Matrix Factorization (NMF). This project separates two sources from a mixed audio signal.

---

📁 Project Structure

.
├── main.py                     # Main script
├── source_separation.py        # NMF dictionary & separation
├── arguments.py                # CLI arguments
├── data_loader.py              # Audio loading
├── utils.py                    # Utilities (SDR, spectrograms)
├── outputs/                    # Output results
└── data/
    ├── train/
    │   ├── source_1.wav
    │   └── source_2.wav
    └── test/
        └── test.wav

---

🚀 How to Run

pip install numpy librosa soundfile scikit-learn matplotlib

python main.py

---

⚙️ Command Line Options

Argument	Description	Default
--data_path	Path to dataset	./data
--output_dir	Directory to save outputs	./outputs
--sr	Sampling rate	48000
--n_components	NMF components	64
--n_fft	FFT window size	1024
--hop_length	Hop size for STFT	512
--eval_sdr	Print SDR before/after	flag
--play_audio	Play result (IPython only)	flag
--save_audio	Save separated audio	flag
--plot_spectrogram	Save spectrogram image	flag

---

📊 Output

The script will save audio and a spectrogram image:

outputs/
├── estimated_source_1.wav
├── estimated_source_2.wav
└── spectrograms.png

---

📚 Reference

Lee, Daniel D., and H. Sebastian Seung.
"Learning the parts of objects by non-negative matrix factorization."
Nature 401.6755 (1999): 788–791.
🔗 DOI: 10.1038/44565