Reproduce “Mel Spectrogram Inversion with Stable Pitch” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bruno Di Giorgi, Mark Levy, Richard Sharp
Apple researchers address pitch instability in neural music synthesis by predicting magnitude spectra and phase gradients, then reconstructing full phase through a dedicated integration algorithm. Their method improves pitch stability in synthesized music, achieving a 60% reduction in harmonic error for single notes compared to existing vocoders.
Key insights
- Speech-optimized vocoders produce pitch instability when applied to musical signals due to their inability to maintain horizontal phase coherence for sustained notes.
- Phase gradients (time and frequency derivatives of phase) exhibit shift-invariant properties for sustained notes, making them suitable targets for neural networks to predict.
- Combining neural network prediction of magnitude and phase gradients with a classification-based phase integration algorithm effectively reconstructs coherent phase information for musical content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.