Reproduce “When De-noising Hurts: A Systematic Study of Speech Enhancement Effects on Modern Medical ASR Systems” 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
Sujal Chondhekar, Vasanth Murukuri, Rushabh Vasani, Sanika Goyal, Rajshree Badami, Anushree Rana, Sanjana SN, Karthik Pandia and 3 more
A systematic evaluation revealed that applying MetricGAN-plus-voicebank speech enhancement consistently degraded the accuracy of modern ASR models when transcribing noisy medical audio, observing up to a 46.6% increase in semantic word error rate. This challenges the traditional assumption that denoising universally improves ASR performance.
Key insights
- Modern, large-scale ASR models, trained on diverse and noisy real-world data, may possess inherent noise robustness, making external speech enhancement redundant or even detrimental.
- Speech enhancement techniques optimized for human perceptual quality can introduce artifacts (e.g., spectral smearing, temporal discontinuities) that are not beneficial and can actively degrade the performance of neural ASR systems.
- The findings suggest a misalignment between the objectives of traditional speech enhancement (perceptual quality) and the learning objectives and internal mechanisms of modern, end-to-end ASR systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.