DeepFilterNet

Why we included this project

DeepFilterNet tackles a specific, common problem: cleaning up speech recorded at 48kHz while keeping the compute low enough for real-time use. Its deep-filtering framework is built with embedded devices in mind, and the project offers more than one way to put it to work. A pre-compiled deep-filter binary strips noise out of WAV files in batch, and a LADSPA plugin lets you wire it into PipeWire as a virtual noise-suppression microphone for calls and streaming. The same codebase includes Python training and evaluation code, so you are not stuck with the pretrained models if you want to fine-tune for your own audio. Between the light footprint and the range of deployment paths, it is a sensible first stop for speech-enhancement experiments and real audio pipelines.