Reproduce “Reverberation as Supervision for Speech Separation” 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
Rohith Aralikatti, Christoph Boeddeker, Gordon Wichern, Aswin Shanmugam Subramanian, Jonathan Le Roux
Aralikatti et al. at Mitsubishi Electric Research Laboratories (MERL) introduce Reverberation as Supervision (RAS), a novel unsupervised loss function that leverages the inherent reverberant differences between two-channel audio to train single-channel speech separation networks. This semi-supervised approach yields a 1.5-1.9 dB improvement in SI-SDR over limited supervised baselines and achieves 70-78% of fully supervised performance using only 5-10% of the labeled training data.
Key insights
- Oracle experiments demonstrated a 5-6 dB SI-SDR improvement when predicting the right-channel mixture from separated left-channel sources compared to the mixed signal, confirming the objective's strong signal for separation.
- The use of independent Wiener filter estimation for each source estimate is critical for stable gradients and effective separation, yielding significantly better performance than joint filter estimation.
- The RAS objective appears to implicitly encourage dereverberation alongside source separation, as higher performance was observed in oracle predictions when using dry or direct-path source signals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.