Most engineers set a frequency, lower the threshold, and call it done. The de-esser is a black box: goes in sibilant, comes out tamed. But there are three fundamentally different machines hiding behind that label, and confusing them explains a class of mix artifacts that most engineers attribute to wrong settings rather than wrong tool selection.
The Wideband De-Esser: A Compressor With a Specific Trigger
The oldest and simplest architecture. Here is what actually happens:
A highpass or bandpass filter sits in the sidechain detection path, tuned to the sibilance range (more on that frequency range below). When signal in that band exceeds the threshold, the compressor fires. The gain reduction lands on the full signal, all frequencies simultaneously.
This means every time a vocalist produces an "s," everything below the detection band also drops. The low mids, the fundamental, the body of the voice. For a brief window, the entire channel loses level. On a solo vocal track with careful settings this may be inaudible. On a vocal riding above a dense arrangement, or on a mix bus, it is audible as a subtle pumping: the low end briefly ducks in sync with consonants.
Technically, a wideband de-esser is a frequency-triggered VCA compressor. Nothing more. The Waves Renaissance DeEsser operates this way, and it remains the dominant architecture in budget-tier tools.
When to use it: light de-essing on isolated vocal tracks, where the gain reduction is small enough that the full-signal attenuation is inaudible. Below about 2-3 dB of peak reduction, the difference between wideband and split-band is hard to hear.
Split-Band: Surgery on One Frequency Region
The logical fix for the pumping problem is to split the signal before applying gain reduction. A crossover filter divides the output into a low band and a high band. The gain reduction only lands on the high band. The low-frequency content passes through entirely unaffected.
Detection still uses a sidechain filter on the full input signal, but the gain reduction element sits only in the high-frequency path. When sibilance fires, the highs compress while the lows continue uninterrupted.
FabFilter Pro-DS in Split Band mode works this way, using linear-phase crossover filters. Linear phase matters here: minimum-phase crossovers introduce phase rotation that varies with frequency, which creates audible coloration when the two bands are recombined after processing. Linear-phase crossovers maintain phase coherence across the crossover point at the cost of latency (FabFilter adds approximately 15 ms with their implementation).
The crossover frequency is typically set close to or slightly below the detection frequency. If you are sweeping for sibilance at 7 kHz, the crossover might sit around 5-6 kHz, preserving the mid-high presence while isolating the sibilant energy above.
Sonnox Oxford SuprEsser uses a similar architecture, combining linear-phase crossover filters derived from their mastering EQ with compressor circuitry from Oxford Dynamics. The SuprEsser adds Auto Level Tracking, which adjusts the threshold relative to the instantaneous signal level, keeping the gain reduction proportional regardless of how loud or soft the vocalist is singing. This is not FFT processing but it is a meaningfully smarter detection stage than pure fixed-threshold comparison.
Split-band de-essers are the correct tool for most vocal de-essing tasks. The high-frequency-only attenuation avoids the full-signal pumping, and the linear-phase crossover eliminates phase smearing on recombination.
Spectral De-Essers: Per-Bin Attenuation
The third architecture is categorically different. iZotope RX's Spectral De-ess does not use crossover filters or VCA-style compression. It runs a short-time FFT analysis on the input, producing a frequency-domain representation of the signal frame by frame. Gain reduction decisions are made individually for each frequency bin.
In Spectral mode (as opposed to Classic mode, which reverts to a broadband gain envelope), only the bins that contain excessive sibilant energy are attenuated. Bins carrying useful high-frequency content such as air, harmonic overtones, consonant definition sit beneath the threshold and pass through at full level.
The practical difference is audible on problem vocals where conventional de-essing removes transparency. A split-band de-esser compresses everything above 5 kHz when an "s" fires. A spectral de-esser might only attenuate a cluster of bins between 7 and 9 kHz where the actual sibilant peak lives. The result on an open, airy vocal can be significantly more natural.
The trade-off is processing latency and CPU overhead inherent to frame-by-frame FFT analysis. This makes spectral de-essers less suitable for zero-latency live monitoring or in-session tracking. They are most appropriate in post-production and mixing, where latency compensation handles the delay.
Release Time and the Distortion Trap
Regardless of architecture, release time is the most common source of de-essing artifacts. When the release is set too short, the gain reduction snaps back before the sibilant transient has fully decayed. The compressed state transitions abruptly to the uncompressed state mid-consonant. To the ear this sounds like distortion or clipping on the tail of the "s" sound. A chopped, crunchy texture rather than smooth attenuation.
A starting point: set release to at least the duration of a typical sibilant event. Sibilant consonants in vocal recordings typically last 40 to 80 ms. A release of 50 to 150 ms covers this range. Too long and the de-esser does not recover before the next transient, leaving consecutive consonants compressed. Sweep from 80 ms down until the distortion appears, then back up slightly.
Frequency Ranges: Where to Start Sweeping
Male voices: sibilant energy typically peaks between 5 and 8 kHz, with 6 kHz as a common starting point for the sweep.
Female voices: typically 6 to 10 kHz, with a peak often around 8 kHz. The female formant structure shifts sibilant energy upward by roughly 2 to 3 kHz compared to male voices.
Neither of these is a fixed target. The actual peak depends on the individual voice, the microphone's high-frequency response, and the consonant being produced. Sweep in narrow Q while the vocalist holds an "s" sound and listen for the frequency that sounds most like the problematic sibilance on the recording.
De-Esser vs. Dynamic EQ: The Real Difference
Split-band de-essing is mechanically close to a single-band dynamic EQ with an internal sidechain filter tuned to the target range. You can build a functional de-esser inside most parametric EQ plugins that have dynamic sections.
The distinction that matters is detection intelligence. FabFilter Pro-DS in Single Vocal mode analyzes the spectral character of incoming signal events, not just their level. It distinguishes between a loud transient that happens to have high-frequency content and a genuine sibilant event. This prevents false triggering on cymbal bleed, bright consonants that are not "s" sounds, or high-frequency room reflections. A conventional dynamic EQ responds only to level exceeding a threshold in a band: no sibilance-specific logic.
For most mixing tasks, a dynamic EQ set to a narrow band with an internal sidechain filter delivers equivalent results to a split-band de-esser. For transparent results on problem vocals with complex content around the sibilance range, a spectral de-esser or an intelligent detection algorithm saves manual tweaking.
Choosing the Right Tool
If you hear low-mid pumping in sync with consonants, you are using a wideband de-esser. Switch to split-band.
If split-band reduces the sibilance but also dulls the air and presence of the vocal even at conservative settings, try a spectral de-esser. The difference is that split-band attenuates everything above the crossover when gain reduction fires. Spectral attenuation leaves the non-sibilant high-frequency content untouched.
On a mix bus or mastering chain, only split-band or spectral architectures are appropriate. Wideband de-essing on a full mix will pump the bass and low mids in sync with every vocal consonant in the arrangement.
The choice of architecture is not a preference. Each design makes a different trade-off between transparency, latency, and processing complexity. Knowing which one is running inside your plugin is the first step to diagnosing why it sounds the way it does.