Reproduce “Adaptive Noise Resilient Keyword Spotting Using One-Shot Learning” 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
Luciano Sebastian Martinez-Rau, Quynh Nguyen Phuong Vu, Yuxuan Zhang, Bengt Oelmann, Sebastian Bader
Martinez-Rau et al. propose a lightweight, continuous noise adaptation method for Keyword Spotting (KWS) systems, leveraging one-shot learning and a single training epoch to fine-tune pre-trained models. This approach significantly enhances KWS accuracy on resource-constrained devices, achieving up to 46.0% improvement in accuracy for specific on-site noise conditions at lower SNRs.
Key insights
- Training with general noisy data (`noise-aware-model`) enhances generalization to known types of noise, outperforming the clean-trained `baseline-model` in such conditions.
- The extremely lightweight 1-shot, 1-epoch fine-tuning successfully enables pre-trained models to adapt to novel, specific on-site noise sources, providing substantial accuracy gains.
- The effectiveness of the adaptation is sensitive to the Signal-to-Noise Ratio (SNR) during adaptation, with lower adaptation SNRs yielding better performance in highly noisy operational conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.