RNNoise

RNNoise is the quiet workhorse of consumer noise suppression. Rather than estimating hundreds of spectral values with a giant network like the 2012-era Google systems did (10M weights and far too heavy for a Raspberry Pi), it splits the spectrum into 22 psychoacoustic Bark bands borrowed from the Opus codec and trains a small recurrent network to predict a per-band gain, while classical pitch filtering handles the voiced harmonics. The result fits in roughly 100 KB total, runs real-time on a Cortex-A53, and sounds noticeably better than the speexdsp MMSE method it was designed to replace. Jean-Marc Valin published the model and code under BSD-3 in 2017 after Mozilla's 'donate your noise' data drive. Adoption is the story here: it is the optional noise filter in OBS Studio, ships in PipeWire and Jitsi's WASM builds, powers NoiseTorch, and sits behind Discord's Krisp-adjacent suppression. The tradeoff is that it's a 2017-era model — it can sound watery or eat quiet speech in conditions far outside its training, and newer models beat it on quality. But nothing this small and this free touches it for deployment footprint.

• Author-reported: significantly better noise reduction than speexdsp MMSE on background-speech, car, and street noise tests
• No independent benchmark replication found (ubiquitous via integration, not benchmarks)