Reproduce “DialogueSidon: Recovering Full-Duplex Dialogue Tracks from In-the-Wild Dialogue Audio” 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
Wataru Nakata, Yuki Saito, Kazuki Yamauchi, Emiru Tsunoo, Hiroshi Saruwatari
DialogueSidon jointly restores and separates degraded monaural two-speaker dialogue audio into clean, distinct speaker tracks. This enables the creation of large-scale, high-quality full-duplex conversational datasets from widely available, in-the-wild recordings to support advanced spoken dialogue research.
Key insights
- Jointly restoring and separating degraded conversational speech is more effective than sequential processing, which often distorts overlapping dialogue or is not robust to heavy degradation.
- Utilizing a compact, semantically rich latent space derived from self-supervised learning features (w2v-BERT) and a VAE enables efficient and robust diffusion-based modeling of speech.
- A diffusion transformer (DiT) can effectively refine speaker-wise latent representations by leveraging permutation-invariant auxiliary predictions and conditioning on the degraded mixture, providing a robust solution for complex dialogue scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.