DeepFilterNet

DeepFilterNet is the denoiser people reach for when the budget is measured in milliwatts. It works in two stages: the first applies coarse gains on a perceptual ERB filter bank to clean up the spectral envelope, and the second performs deep filtering — small learned complex FIR filters applied per time-frequency bin that reconstruct periodic speech structure where simple masking fails. Because the feature space is compressed to ERB bands and the network uses depthwise separable convolutions, the whole thing is about 2 million parameters and runs in real time on constrained hardware at full 48 kHz bandwidth. Author-reported results on the VCTK/DEMAND set put it ahead of PercepNet and DCCRN at a fraction of the cost, with the deep-filtering stage alone adding roughly 2.8 dB SI-SDR and 0.24 PESQ. It also handles dereverberation and has a training target clamp so devices like hearing aids can keep some environmental awareness. The repo is actively maintained with several pretrained checkpoints (DNS4-trained), though the Python/Rust packaging has gone stale enough that newer torchaudio versions can break the install.

• Paper (ICASSP 2022/2023, author-reported): outperforms PercepNet and DCCRN on VCTK/DEMAND at lower compute; deep filtering adds ~2.8 dB SI-SDR, ~0.24 PESQ over ERB gains alone
• No independent benchmark replication found