Reproduce “Interaural time difference loss for binaural target sound extraction” 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
Carlos Hernandez-Olivan, Marc Delcroix, Tsubasa Ochiai, Naohiro Tawara, Tomohiro Nakatani, Shoko Araki
Researchers at NTT Corporation introduced a differentiable Interaural Time Difference (ITD) loss to improve spatial cue preservation in deep learning-based binaural Target Sound Extraction systems. This novel approach, which optimizes cross-correlation coefficients, led to superior performance across all measured spatial metrics, including a 16% relative reduction in ITD errors, without degrading sound extraction quality.
Key insights
- Explicitly incorporating spatial information into the training loss of binaural TSE systems is more effective for preserving spatial cues than relying solely on implicit learning from signal-level losses.
- Optimizing for Interaural Level Difference (ILD) alone is insufficient for comprehensive spatial cue preservation, and while IPD loss reduces phase errors, it does not reliably improve ITD errors.
- The proposed differentiable ITD loss, based on cross-correlation, effectively overcomes the non-differentiability challenge and provides superior preservation of all binaural spatial cues.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.