what a dynamic resonance suppressor is

the problem it solves

record a vocal, a trumpet, or an acoustic guitar and you will hear it: a frequency that sounds fine when the note is not there, and wrong when it is. the body of the room amplifies 240 Hz. the singer’s chest resonates at 380 Hz on certain vowels. the acoustic guitar has a wolf note at 480 Hz that jumps three dB above every other pitch.

a static EQ cut at 240 Hz removes the problem when the note is present. it also removes that frequency from every other moment in the performance, including the silence. the note loses body. the room sounds smaller. you removed more than you intended.

the question a dynamic resonance suppressor answers: what if the cut was only there when the problem was there?

what the category actually is

a dynamic resonance suppressor watches the frequency spectrum in real time, identifies peaks that stand out against the surrounding spectral content, and applies gain reduction to each peak, proportional to how much it stands out, only while it is there.

three things distinguish it from adjacent tools:

it is not a static EQ. static EQ applies a fixed gain curve that does not change with the signal. a -3 dB cut at 3.2 kHz removes 3 dB at 3.2 kHz at all times, on every moment of audio that passes through. a resonance suppressor applies that cut only when 3.2 kHz is elevated above its neighbours. when the resonance is absent, the gain curve is flat.

it is not a dynamic EQ. a dynamic EQ monitors signal level at user-selected frequencies. you choose which bands to watch and set thresholds manually. typically 4-6 bands. a resonance suppressor scans the full spectrum automatically. you do not tell it which frequencies to watch. it decides by comparing each frequency against the spectral content around it. the response is proportional to how much each peak stands out, not just whether a level threshold was crossed.

it is not multiband compression. multiband compression divides the spectrum into a small number of fixed ranges and compresses each range when the overall level in that range crosses a threshold. a resonance suppressor is more granular: each band can respond independently, and the criterion is spectral shape (does this frequency peak above its neighbours?) rather than overall loudness in a region.

how detection works

the signal enters the processor and is converted to the frequency domain. in most implementations this is done via fast Fourier transform (FFT). the FFT produces a snapshot of the signal’s magnitude at each frequency at that moment in time.

the processor then builds a reference: a smoothed estimate of the spectral shape, representing what the signal “should” look like without resonances. the smoothing function is the key design choice. smooth it too aggressively and you get a flat floor that treats normal tonal content as resonance. smooth it too little and you miss narrow peaks.

each frequency is compared against the smoothed floor at its location. if it exceeds the floor by more than the threshold (typically user-adjustable as “depth” or “sensitivity”), gain reduction is applied at that frequency. the amount of reduction scales with how far above the floor the peak sits.

the detection and reduction process happens thousands of times per second, tracking resonances that shift in pitch, grow and shrink, and appear on some notes and not others.

how gain reduction is applied

once a peak is detected, the processor applies gain reduction in that frequency region. the critical design question is: how broad is the reduction, and how fast does it move?

bandwidth. the gain reduction should be narrow enough to remove the resonance without pulling down the surrounding frequencies that did not exceed the threshold. too wide and you hollow out the spectrum. too narrow and the resonance leaks around the cut.

time constants. gain reduction is not instantaneous. attack determines how quickly the suppressor responds to a rising peak. release determines how quickly it lets go when the peak subsides. fast attack and fast release can create audible artifacts on transient material. slower time constants trade responsiveness for transparency.

frequency-dependent behaviour. some processors apply the same time constants across the full spectrum. others vary them by frequency: faster in high-frequency bands where resonances are more transient, slower in low-mid bands where notes sustain longer. this is one of the design choices that separates different suppressor architectures.

the linear vs perceptual band question

most resonance suppressors work in linear frequency: each FFT bin covers the same number of Hz. a bin at 200 Hz and a bin at 8000 Hz represent equal frequency slices.

the problem is that your ear does not hear in linear frequency. cochlear resolution is roughly logarithmic: your hearing is most sensitive and most discriminating in the 200 Hz to 5 kHz range, and progressively coarser above and below. a 50 Hz-wide resonance at 250 Hz is a one-critical-band problem. a 50 Hz-wide resonance at 8000 Hz sits inside a much wider critical band and the ear localises it differently.

a suppressor that operates on linear bins applies equal processing weight across the spectrum regardless of where the ear notices the problem. it treats the 8 kHz resonance with the same discrimination as the 300 Hz resonance, even though the ear’s sensitivity to each is very different.

the alternative is to build the detection and reduction on a psychoacoustic scale, most commonly the equivalent rectangular bandwidth (ERB) scale derived by Glasberg and Moore.¹ ERB bands are narrow where ears are sensitive and wider where they are not. a suppressor built on ERB spacing concentrates its detection exactly where you perceive the problem.

this architectural choice does not make a suppressor “better” in absolute terms. it makes the processing match the perceptual target more closely. whether that difference is audible on any specific source depends on the material.

SMOOTH as a worked example

full disclosure: i make KERN SMOOTH.

SMOOTH uses a 4096-point FFT to analyse the incoming signal, then maps that analysis onto 40 ERB-spaced bands derived from the Glasberg and Moore scale. each band has its own detection threshold, attack time, and release time. bands above 5 kHz respond faster than bands below 200 Hz, reflecting the different sustain characteristics of high-frequency resonances versus low-mid room modes.

the signal flow is: STFT analysis, ERB-band mapping, per-band detection against a smoothed spectral reference, per-band gain reduction, STFT resynthesis, output.

M/S routing lets SMOOTH process mid and side independently. a resonance in the mid channel (the centred image: lead vocal, kick, bass) can be suppressed without touching the side channel (the stereo content: reverb tails, panned doubles, room information). this is the architectural decision that most single-suppressor articles leave out.

the 4096-point FFT introduces approximately 93 ms of latency at 44.1 kHz. the DAW compensates automatically during playback and mixing. SMOOTH v1.4.0 adds an opt-in LIVE low-latency mode (about 12 ms) for tracking and monitoring use.

CPU: under 3% on macOS at 44.1 kHz, single stereo instance.

when to reach for a resonance suppressor

a resonance suppressor is the right tool when:

• the problem is a frequency peak that appears dynamically and disappears between notes, phrases, or gestures. static EQ cannot distinguish present from absent.
• the peak is narrow enough that a static cut would hollow out audible tonal content when the resonance is not there.
• the problem shows up on multiple sources simultaneously and manual per-source static EQ work is consuming session time.

it is not the right tool when:

• the problem is consistent broadband harshness: a mix that is too bright overall. a high-shelf or tilt EQ addresses that more transparently.
• the problem is a discrete, predictable frequency that does not change with the performance. a static narrow notch is the more precise and less computational solution.
• the problem is dynamic range or transient control. that is compression or limiting territory.

the lineage

resonance suppression as a dedicated category is largely a software innovation. the combination of full-spectrum automatic detection and per-bin or per-band gain reduction became practical only with cheap FFT computation in plugin form.

oeksound’s original Soothe (2016) is the product that named the category and made it a standard part of the mixing toolkit. before Soothe, similar results were achieved with chains of dynamic EQ bands or by routing through noise reduction plugins in ways their designers had not intended. Soothe turned an improvised technique into a single-purpose tool, and the category has grown substantially since.

the underlying mathematics of spectral suppression is older: the Ephraim-Malah minimum mean-square error spectral estimator from 1984² established many of the temporal-smoothing principles that modern suppressor plugins use to minimise musical-noise artifacts, even if the implementation context has changed.

frequently asked questions

frequently asked questions

what is a dynamic resonance suppressor?

a dynamic resonance suppressor is a spectral processor that detects resonant frequency peaks in a signal and applies gain reduction to those peaks, only when they exceed a threshold, only at the frequencies where they appear. it differs from a static EQ because it is inactive when the resonance is absent. it differs from a dynamic EQ because it monitors the entire frequency spectrum simultaneously rather than a set number of fixed bands.

how is a dynamic resonance suppressor different from a dynamic EQ?

a dynamic EQ applies gain reduction at user-defined frequencies when signal in those bands exceeds a threshold. you choose 4-6 bands and set the thresholds manually. a resonance suppressor scans the full spectrum automatically, finds peaks, and applies suppression to each one independently. you do not choose which frequencies it watches. the suppressor decides based on what the signal is doing at any given moment.

can I do the same thing with multiband compression?

not precisely. multiband compression divides the spectrum into a small number of fixed frequency ranges (typically 4-6) and compresses each range when its overall level exceeds a threshold. resonance suppression is more granular: each analysis band (which may be one of dozens or hundreds) can respond independently, and the response is proportional to how much the peak stands out relative to the surrounding spectrum. multiband compression catches broadband loudness; resonance suppression catches spectral spikes.

what does ERB-domain processing mean?

ERB stands for equivalent rectangular bandwidth: a psychoacoustic scale that models how the cochlea resolves frequency. unlike linear frequency spacing, ERB bands are narrow in the 200 Hz to 5 kHz region where human hearing is most sensitive, and wider in the sub-bass and high air regions where it is less sensitive. a resonance suppressor built on ERB spacing concentrates its detection and correction where you actually perceive the problem.

what is a resonance suppressor used for?

the most common uses are: removing harsh frequency peaks from vocals, acoustic instruments, and bright synths; de-resonating rooms and recorded spaces; taming cymbal and percussion harshness on mix buses; and spectral unmasking by using an external sidechain signal to tell the suppressor which frequencies to remove from a second source. the unifying thread is unwanted spectral peaks that appear and disappear dynamically, which static EQ cannot remove without creating audible artifacts when the resonance is not present.

references

Footnotes

1. Glasberg, B. R. and Moore, B. C. J. (1990). “Derivation of auditory filter shapes from notched-noise data.” Hearing Research, 47(1-2), 103-138. The foundational paper establishing the ERB scale used by psychoacoustic processors. ↩

2. Ephraim, Y. and Malah, D. (1984). “Speech enhancement using a minimum mean-square error short-time spectral amplitude estimator.” IEEE Transactions on Acoustics, Speech, and Signal Processing, 32(6), 1109-1121. The temporal-smoothing framework underlying spectral suppression gain estimation. ↩