Reproduce “DPCRN: Dual-Path Convolution Recurrent Network for Single Channel Speech Enhancement” 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
Xiaohuai Le, Hongsheng Chen, Kai Chen, Jing Lu
A Dual-Path Convolution Recurrent Network (DPCRN) is proposed for single-channel speech enhancement, effectively removing noise and reverberation by integrating 2D convolutional processing with a Dual-Path Recurrent Neural Network for efficient long-sequence modeling. The model achieved competitive performance on the Interspeech 2021 Deep Noise Suppression (DNS) Challenge, including a third-place ranking in the wide band scenario with an overall MOS of 3.57.
Key insights
- The Dual-Path RNN (DPRNN) module effectively addresses long-sequence modeling challenges in the time-frequency domain by processing audio in both intra-chunk (spectral) and inter-chunk (temporal) dimensions.
- Combining a 2D convolutional encoder-decoder with a DPRNN central module allows DPCRN to capture hierarchical spectral-temporal features while maintaining global context.
- The model achieves competitive speech enhancement performance with a relatively small parameter count (0.8 million), indicating computational efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.