> ## Documentation Index > Fetch the complete documentation index at: https://comfytv.org/llms.txt > Use this file to discover all available pages before exploring further. # Noise Reduction (Spectral) > Learn the sound of steady background noise and subtract it, spectrally — like Audacity's Noise Reduction. Noise Reduction (Spectral) ## What this node does **Noise Reduction (Spectral)** removes broadband, steady-state noise (hiss, hum, air-con drone) using a spectral-gate approach: it builds a noise profile and attenuates frequency content that matches it, leaving the wanted signal. You can either let it estimate the noise from the track itself, or hand it a dedicated **noise\_sample** (a snippet of noise-only audio) for a cleaner profile. It accepts a `COMFYTV_AUDIO` snapshot or a `COMFYTV_VIDEO` (its audio track is used; audio takes priority), with an optional second `COMFYTV_AUDIO` for the noise sample. It outputs processed `COMFYTV_AUDIO` and has a ▶ **Run**. To feed native ComfyUI `AUDIO` in or out, insert a **Bridge** — see [bridges.md](https://github.com/jtydhr88/ComfyTV/blob/main/docs/bridges.md). ## When to use it * Clean up hiss or steady room tone under a voice or music recording. * Remove air-conditioner or fan drone captured during a take. * Use a silent-but-noisy passage as a **noise\_sample** to precisely target that noise. ## Parameters ### reduction\_db How much the matched noise is attenuated, in dB. Range **1 to 48**, default **12**. Higher removes more noise but risks artifacts and a hollow, "underwater" sound. ### sensitivity How aggressively frequency content is judged to be noise. Range **0.5 to 24**, default **6**. Higher classifies more as noise (more removed); lower is more conservative and preserves detail. ### freq\_smooth\_bands Smoothing of the noise gate across neighboring frequency bands. Range **0 to 12**, default **6**. More smoothing reduces "musical noise" (tinkling artifacts) at the cost of some sharpness. ### noise\_sample Optional second audio input: a snippet containing only the noise. When provided, the profile is built from it instead of estimated from the main track, usually giving a cleaner result. ## Outputs | Output | Type | Meaning | | --------- | --------------- | --------------------------- | | **audio** | `COMFYTV_AUDIO` | The denoised audio snapshot | ## Tips * Start with **reduction\_db** around 12 and raise only as needed; over-reduction sounds artificial. * If you hear tinkly "musical noise" artifacts, increase **freq\_smooth\_bands**. * A dedicated **noise\_sample** almost always beats self-estimation — grab a second of the room tone before/after the take. ## Related nodes * **Audio Denoise** — FFmpeg-based `afftdn`/`anlmdn` denoisers and a silence trimmer, an alternative approach. * **Audio Stem Split** — isolate the wanted stem, which can be a cleaner starting point than denoising a full mix.