← All posts ·

Frequency Dependent Gating: 5 Setup Steps for Mixes

Frequency Dependent Gating: 5 Setup Steps for Mixes

Engineer adjusting spectral gating during a mix

Frequency-dependent gating is a spectrogram-based gate that selectively mutes time-frequency components, letting you remove noise or shape dynamics in one band without touching the rest of the signal. Tools like Noisereduce use spectral gating to mask out noisy bins based on per-frequency statistics, and our own work on per-band processing at Vector DSP follows the same principle: treat each frequency range as its own problem instead of applying one gate to the whole mix.

TL;DR:

  • When available, profile a few seconds of noise only; otherwise, use nonstationary estimation, then start near a 1.5 to 2.0 standard deviation threshold and 80% reduction.
  • Heavy gating can leave musical noise or remove wanted harmonics; when noise overlaps the target or the signal is very weak, consider machine learning denoising.
  • For tracking, favor a low latency multiband gate with lookahead under a couple of milliseconds; reserve finer per bin processing for offline cleanup.

Table of Contents

How frequency-dependent gating works in the time-frequency domain

Traditional gating looks at one signal: above the threshold, it passes; below it, it closes. Frequency-dependent gating works on a spectrogram instead, the output of a short-time Fourier transform (STFT) that breaks audio into overlapping frames and converts each into a set of frequency bins over time. Instead of one gate, you get a grid of decisions: for every bin at every moment, is this signal or noise?

That grid becomes a time-frequency mask, a set of multipliers applied to the magnitude spectrum before the audio is reconstructed with an inverse STFT. Coarse implementations group bins into bands, similar to a multiband compressor, and gate each band with its own threshold. Finer implementations work per-bin, producing a mask with as much resolution as the FFT size allows.

STFT frames forming a time-frequency gate mask

Masking and spectral subtraction solve the same problem differently. A mask scales each bin down toward zero when it is judged to be noise, which tends to sound smoother but can still leave a residual texture. Spectral subtraction estimates the noise spectrum and subtracts it directly from the signal spectrum, which can remove noise more aggressively but is prone to "musical noise," a scattering of isolated tonal artifacts at random bins. This is why most modern tools smooth the mask before applying it: a Gaussian filter across frequency and time spreads the gating decision over neighboring bins and frames rather than letting isolated pixels snap open or shut, which keeps structured harmonic content from being chopped into artifacts.

Algorithms you'll encounter: Noisereduce, 2D spectral gating, and spectral subtraction

Most spectral noise tools build on one of a few core algorithms, and knowing which one a plugin uses tells you what to expect from it.

Noisereduce-style spectral gating estimates a mean and standard deviation for each frequency bin, then masks out any bin that falls below a z-score threshold relative to that bin's noise statistics. It runs in stationary mode, using a fixed noise profile, or nonstationary mode, which updates its estimate on a sliding window as the input changes. Smoothing parameters such as frequency-mask smoothing and time smoothing control how much the mask blurs across neighboring bins and frames before it is applied, which the Noisereduce research credits with reducing the harsh, binary on/off feel of early gating approaches.

2D spectral gating extends this by applying a Gaussian filter across both time and frequency before thresholding, which better preserves signals that spread energy across multiple frequencies at once, like a buzzing fan or a sustained vocal harmonic, rather than treating each bin in isolation.

Spectral subtraction, the older of the family, estimates a noise spectrum and subtracts it outright. Evaluations of short-time spectral attenuation for musical restoration found this approach effective but sensitive to window length, since overly short analysis frames introduce distortion in tonal material. None of these methods handle highly nonstationary noise well: when noise shifts quickly or closely resembles the target signal, supervised machine learning models can outperform simple gating, though they require labeled training data and far more compute.

Algorithms you'll encounter: Noisereduce, 2D spectral gating, and spectral subtraction — overview diagram

When to reach for frequency-dependent gating in a mix

Frequency-dependent gating earns its place when a problem lives in a specific part of the spectrum rather than across the whole signal.

It is a poor fit when the signal-to-noise ratio is extremely low or when the noise and the target signal overlap heavily in both time and frequency. In those cases, gating tends to either leave noise behind or eat into the wanted signal, and a cleaner recording or a different noise-reduction approach serves you better.

A practical workflow with starting settings you can copy

Getting a usable result quickly comes down to picking sane starting values and adjusting from there rather than guessing from scratch.

  1. Set your analysis window. Use a longer FFT size, around 2048 to 4096 samples, for vocals and sustained tonal material where frequency resolution matters more than speed; drop to around 1024 for drums or anything transient-heavy where time resolution matters more. Evaluations of musical restoration point to analysis windows around 40 to 50 milliseconds as a reasonable middle ground for tonal material, since shorter windows risk canceling low-level sinusoids.
  2. Choose your noise reference. Supply a short noise-only clip when one is available (a few seconds of room tone, for instance), or switch to nonstationary mode, which updates its noise estimate continuously as it processes.
  3. Set your threshold and reduction amount. Start around a 1.5 to 2.0 standard-deviation threshold with a prop_decrease value near 0.8, which attenuates identified noise by 80% rather than muting it entirely, a setting the Noisereduce documentation uses as a conservative default.
  4. Dial in smoothing. Start with modest frequency and time smoothing, widening the Gaussian window if you hear choppy or stuttering artifacts on sustained notes, and narrowing it if transients start to smear.
  5. Verify by ear and by eye. Solo the processed track, bypass repeatedly, and apply makeup gain to match loudness before judging the result. Check a spectrogram view if your editor offers one, such as Audacity online, to confirm you're not carving into wanted harmonics.

Pro tip: Always compare at matched loudness; gated audio often sounds "cleaner" purely because it's quieter, not because the noise is actually gone.

Real-time and plugin considerations for live and offline use

Spectral gating behaves differently depending on whether it runs live in a DAW or offline on a finished file. Lookahead is the key constraint: keeping it under a couple of milliseconds matters for tracking and monitoring, where any added delay becomes audible as a performer plays, while offline cleanup and mastering can afford a longer lookahead window in exchange for cleaner analysis.

CPU cost scales with resolution. Per-bin spectral masks, which analyze hundreds of frequency bins independently, cost more to compute than coarse multiband gates that group the spectrum into a handful of wide bands, so plugins built for real-time tracking tend to favor the coarser approach, saving per-bin precision for offline passes. Format support also matters in practice: a plugin built on VST3, AU, and AAX will cover most DAW setups, and automation lanes become useful once you're riding threshold or reduction amount across a song section. As a general rule, keep a low-latency multiband insert in the mix during tracking and recording, and save finer per-bin spectral cleanup for an offline pass once the take is set. Our guide to avoiding aliasing in plugins covers related tradeoffs in choosing analysis resolution without introducing artifacts.

Why predictable gating behavior comes down to deliberate filter design

Gating only feels trustworthy when it behaves the same way on every source and every sample rate. That consistency comes from deliberate filter design: proportional Q behavior, stable filter coefficients, and smoothing math that doesn't shift character as you change settings. We write more about these choices in our post on frequency-selective effects and proportional Q, and in how we approach frequency-selective reverb for similar reasons: minimal latency and stable behavior matter more than flashy presets once you're relying on a tool daily.

— Kai

How ToneLab handles per-band frequency control

ToneLab is built around a multi-lane parallel effects architecture with per-lane EQ targeting, so instead of reaching for a separate gate plugin per frequency range, you route and shape bands within a single insert. It runs on real-time, low-latency DSP and ships in common plugin formats, so it drops into the same session as anything else on your track.

Vector-dsp

What you need How ToneLab fits
Gate one band without affecting others Per-lane EQ targeting inside one insert
Keep latency low while tracking Real-time, low-latency DSP architecture
Use it in your existing DAW VST3, AU, and AAX support

If frequency-specific control over noise or dynamics is a recurring task in your sessions, ToneLab is priced at $73.99, one-time license, with a demo and full product overview available before you decide.

FAQ

What's the difference between frequency-dependent gating and a regular noise gate?

A regular noise gate opens or closes based on the full signal's level, affecting every frequency equally. Frequency-dependent gating analyzes a spectrogram and gates individual bands or bins independently, so you can remove noise in one frequency range while leaving the rest of the signal untouched.

What does n_fft actually control in spectral gating?

The n_fft parameter sets the size of the Fourier transform window, which determines frequency resolution: larger values like 2048 or 4096 give finer frequency detail and suit sustained, tonal material. Smaller values like 1024 trade frequency precision for better time resolution, which helps preserve transients like drum hits.

Can frequency-dependent gating damage a recording if I push it too hard?

Pushed too aggressively, spectral gating can produce "musical noise," scattered tonal artifacts where bins are masked inconsistently, or it can thin out wanted harmonic content along with the noise. Starting with a moderate prop_decrease value and checking results at matched loudness helps avoid over-processing.

Is frequency-dependent gating useful in real-time tracking, or only for mixing after the fact?

It works in both contexts, but the approach differs: live tracking favors coarse, low-latency multiband gating with minimal lookahead, while offline cleanup can use finer per-bin analysis with more aggressive smoothing since latency isn't a constraint. Many engineers use a lighter real-time version while tracking and a more detailed offline pass during cleanup or mastering.

Do I need machine learning to get good results, or is spectral gating enough?

Spectral gating handles most common studio noise problems, like hum, hiss, and bleed, without any training data or added compute cost. Machine learning-based denoising can outperform it in very low signal-to-noise situations or when noise closely overlaps the target signal, but it requires labeled data and significantly more processing power.

Sources

Recommended