Audio Denoising – 60 Hz Hum-Removal

code, results and an interactive demo for Spectrogram U-Net that suppresses mains hum in music recordings.

---

Project overview

This repo trains a U-Net that predicts a soft mask over magnitude spectrograms and removes the hum while preserving the insturment sound

• Input: single-channel 16 kHz WAV
• Output: denoised WAV (same length)
• Training target: SI-SDR ↑ / PSNR ↑

---

Dataset

Split	Files	Duration
train	2 621	≈ 72 h
val	562	≈ 15 h
test	188	≈ 5 h

Each example contains:

1. Clean music excerpts (from Neo Scholars Take Home Google Drive)
2. Synthetic 60 Hz + harmonics + random phase + slowly-varying amplitude
3. Weighted mix at SNR ∈ {0, 5, 10, 15} dB

Scripts live in dataset/:

# Download music, synthesise hum, build CSV metadata
python dataset/build_dataset.py --root dataset

---

Model architecture

Spectrogram U-Net (Oh, J., Kim, D., & Yun, S.-Y. (2018)) operating on STFT magnitudes (513 bins):

                            +------+               +------+
       │ mag │              | up 4 |<--skip--+-----| out  |
  x ──▶│(B,1)│──inc──down1──| ...  |        |     +------+
       +------+             +------+        |
                         ···                        sigmoid→ soft mask ∈ [0,1]

• Base channels = 32 (≈ 1.7 M params)
• Bilinear up-sampling, sigmoid on logits → mask
• Estimate complex STFT: mask × noisy STFT, then iSTFT

Hyper-parameters (see train_spectrogram_unet.py):

argument	value	note
--epochs	50	cosine LR decay
--lr	3 × 10⁻⁴	AdamW
--batch	16	1-s crops (44 100 samples)
Loss	0.7 · L1(spec) + 0.3 · L1(time)
Aug.	random crops, amplitude normalisation

---

Results

Training vs. validation loss and SI-SDR stay closely aligned, indicating the model converges without over-fitting.

method,split,si_sdr [dB],psnr [dB]
identity,test,7.50,44.24
notch,test,7.04,43.93
unet (best.pt),test,26.30,62.79

• +18.8 dB SI-SDR over raw noisy input ⇒ ~75× noise power reduction
• Fixed notch filter barely helps because music energy also sits at 60 Hz and harmonics.

Generate CSV & figures:

python scripts/evaluate_models.py \
  --dataset dataset --split test --checkpoint checkpoints/best.pt \
  --out results/results.csv
python scripts/plot_results.py results/results.csv

🔊 Listen

noisy	denoised
sax – noisy 🎧	sax – denoised 🎧

---

Quickstart

# create env (conda or venv) and install deps
pip install -r requirements.txt

#  Train (≈ 5 h on one RTX 3060, CPU works but slower)
python train_spectrogram_unet.py --dataset dataset --epochs 50 --batch 16

#   Evaluate best checkpoint
python scripts/evaluate_models.py --dataset dataset --split test \
       --checkpoint checkpoints/best.pt --out results/results.csv

---

Local Gradio app

python scripts/denoise_demo.py

Then open http://127.0.0.1:7860 and drag-&-drop a WAV.

---

Reproducibility

1. Set random seeds in train_spectrogram_unet.py (disabled by default).
2. All checkpoints, CSVs and figures are version-controlled – see checkpoints/best.pt.

References

• Oh, J., Kim, D., & Yun, S.-Y. (2018). Spectrogram-channels U-Net: A source separation model viewing each channel as the spectrogram of each source. arXiv preprint arXiv:1810.11520.

• Zhang, Y., & Li, J. (2023). BirdSoundsDenoising: Deep Visual Audio Denoising for Bird Sounds. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 3164–3173.

• Babaev, N., Tamogashev, K., Saginbaev, A., Shchekotov, I., Bae, H., Sung, H., Lee, W., Cho, H.-Y., & Andreev, P. (2024). FINALLY: Fast and Universal Speech Enhancement with Studio-like Quality. Advances in Neural Information Processing Systems (NeurIPS).

• Zhang, Y., & Li, J. (2023). Complex Image Generation SwinTransformer Network for Audio Denoising. Proceedings of INTERSPEECH 2023, Dublin, Ireland.

---

Made with ❤️ & PyTorch 2.7 ✨