spectral gating

Spectral gating is a noise reduction method that processes audio by converting it into the frequency domain, then silencing or attenuating components below a threshold in each frequency band independently. Unlike conventional gating, which works on amplitude across the entire signal, spectral gating operates on individual frequency bins, making it far more selective. A 1 kHz tone buried in broadband hiss can be preserved while the hiss itself is removed, since the gate respects frequency rather than treating the signal as a single entity.

The process begins with a short-time Fourier transform (STFT) or similar spectral decomposition, typically using window sizes of 512 to 4096 samples depending on frequency resolution requirements. Each frequency bin receives its own threshold, usually estimated from a noise profile captured during a quiet section of the recording. When the magnitude of a bin falls below its threshold, the gate mutes or reduces that component by a fixed amount, often 6 to 20 dB. The gated spectrum is then converted back to the time domain using an inverse STFT, reconstructing audio with suppressed noise.

Threshold setting determines the trade-off between noise reduction and signal degradation. Too aggressive a gate removes wanted content and produces spectral holes, where legitimate signal components disappear and leave audible artifacts or a hollow timbre. Too lenient a gate leaves audible noise. Many implementations use a soft gate rather than hard muting, allowing a gentler roll-off below threshold to avoid sudden frequency discontinuities. Overlap-add methods during reconstruction help maintain phase coherence and smooth transitions.

Practical constraints and variations

Spectral gating works well on stationary noise (tape hiss, electrical hum, air conditioning background) because the noise profile remains stable across time. It struggles with non-stationary noise (traffic, crowd chatter, door slams) unless the gate adapts continuously. Some systems use voice activity detection or other classifiers to update the noise estimate only during confirmed silent frames. The computational load is moderate: an STFT-based gate running at 48 kHz with 2048-sample windows requires roughly 50 to 100 floating-point operations per sample on modern hardware.

In audio mastering and restoration, spectral gating competes with spectral subtraction and Wiener filtering. Spectral gating has the advantage of binary or near-binary decision-making (a bin is either gated or passes through), making it interpretable and predictable. It appears in noise reduction plugins, restoration software for archival recordings, and broadcast audio chains. Creative sound design exploits spectral gating to emphasize transients or carve out frequency content for artistic effect, though it is less common in that role than multiband compression.