bin2bin Music Signals Inpainting

This page contains some listening samples to evaluate the performance of bin2bin, a cGAN-based spectrogram inpainting approach, described in: A Score-aware Generative Approach for Music Signals Inpainting - C. Aironi, L. Gabrielli, S. Cornell, S. Squartini (2023)

Architecture overview

The bin2bin architecture is composed of a U-Net for CQT spectrogram inpainting (G). Multi-objective loss for training the U-Net is obtained by ensembling a PatchGAN discriminator BCEloss, along with a Spectral Convergence loss L_SC, and the MSE between the true, input-aligned piano roll, and that obtained from a pitch estimation module.

Small gap (375 ms)

---

---

---

Medium gap (750 ms)

---

---