Reproduce “STFT-Domain Neural Speech Enhancement with Very Low Algorithmic Latency” 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

Zhong-Qiu Wang, Gordon Wichern, Shinji Watanabe, Jonathan Le Roux

The system reduces algorithmic latency in speech enhancement by analyzing long input windows while overlap-adding only short output tails, with an option to predict one frame ahead. It combines frame-online neural networks with frequency-domain beamforming for microphone arrays. In simulated six-microphone enhancement, it reports higher scores than multi-channel Conv-TasNet at 2 ms, with similar computation; processing time was not measured.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.