Reproduce “Lightweight DNN for Full-Band Speech Denoising on Mobile Devices: Exploiting Long and Short Temporal Patterns” 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
Konstantinos Drossos, Mikko Heikkinen, Paschalis Tsiaflakis
A lightweight Deep Neural Network developed by Nokia Technologies and Nokia Bell Labs enables causal, low-latency full-band speech denoising optimized for mobile devices, achieving an SI-SDR of 22.34 dB with a real-time factor below 0.02 and utilizing 100 times fewer Multiply-Accumulate Operations than previous methods.
Key insights
- Achieving superior signal reconstruction quality (SI-SDR) in full-band speech denoising is possible even with strict causality and ultra-low computational resources through careful architecture design.
- Explicitly modeling both short and long temporal patterns within a lightweight, causal framework significantly enhances denoising performance without compromising real-time constraints.
- Learned dimensionality reduction mappings at the input and output stages are critical for handling the high-frequency information of full-band audio efficiently within a resource-constrained neural network.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.