RNNoise is a compact, open-source recurrent neural network for real-time noise suppression, designed to be light enough to run frame by frame on an ordinary CPU. It combines classic signal-processing structure with a small RNN that estimates how much of each frequency band is speech versus noise, then attenuates the noise while preserving the voice. Its efficiency and permissive licence made it a common baseline in open-source voice stacks and a popular building block for plugins and embedded systems. It won't match the heaviest commercial or GPU models on extreme noise, but for the cost — a few percent of a CPU core — it removes a remarkable amount of steady and intermittent background sound.
The book · Volume 4 of 7
Audio for Video: A Complete Guide to Sound in Video Products: From PCM and Codecs to Loudness, Accessibility, and Real-Time Voice Processing
RNNoise is a popular free baseline for noise suppression because it runs on any CPU, but it has limits on hard noise and music. This volume covers RNNoise in detail with classical suppression, Krisp and other neural models, and the AEC, AGC and VAD stages around it, so you can decide when it is enough. Written by Nikolay Sapunov, CEO at Fora Soft.
Deciding whether RNNoise is good enough for noise suppression in your calling app?
Share noisy recordings from your users and we will benchmark RNNoise against alternatives and recommend a setup with CPU and quality numbers. 250+ video products shipped since 2005.
