Noise suppression removes the steady and intermittent background sounds around a talker — fans, traffic, keyboard taps, a barking dog — while leaving speech intact and natural. Classic approaches estimate the noise spectrum during pauses and subtract it, which works for stationary hums but smears speech and struggles with sudden sounds. Modern systems are deep neural networks (RNNoise, Krisp, NVIDIA RTX Voice) trained on huge mixes of speech and noise; they separate voice from everything else far more aggressively, even pulling a clear talker out of a noisy cafe. The tradeoffs are CPU or GPU cost and the risk of over-suppression producing a processed, underwater quality, so tuning matters.
