Chamfer De-Esser

The threshold is relative and level-tracked, not a line you chase. Turn the singer up 18 dB or down 18 dB and the delivered depth moves by 0.5 dB or less at the command level, and 0.75 dB or less measured end to end, so a threshold set on the verse is still right in the chorus. Each ess is handled on its own and reshaped spectrally rather than ducked: identically constructed events come out identical, with residual below -120 dBFS on every style and mode. At matched delivered depth on the same takes, held-vowel body moves 0.00 dB on average, and a frozen 10 dB floor on ess over body is enforced where gain is applied, so Chamfer cannot be driven into a lisp. The display is the instrument. The pale curve is your input, the lit curve is what leaves, and the shade between them inside the band is exactly the energy that came out; the input tap is latency-aligned to the output tap, so that shade cannot show reduction that did not happen. The band pane is where detection listens: drag its edges, drag the pane to move both and keep the width, scroll to widen or narrow around its centre, or type exact edge frequencies. The RIDE line across it is the detector's own tracked reference, K-weighted over about 250 ms, and esses visibly cross it when they trigger. Under the glass, the ess strip records every detected event, its height the depth applied and its brightness the confidence. Click one and the card freezes that ess: its spectrum before and after, the removed energy shaded, the depth, the confidence and a marker at its centre frequency. AMOUNT sets how far each ess is reshaped toward the voice around it, 0 to 150 percent, and RANGE is the hard cap on how far it may go, 0 to 24 dB. STYLE picks the shape of the cut: Smooth tilts the whole band, Split is the classic band cut, and Surgical reshapes each ess bin by bin toward the target. All three are gain shapes on one engine, so a style change glides and never relatches. SPEED trims the automatic ballistics by up to 50 percent either way, OUTPUT is a plus or minus 12 dB trim, and BYPASS is gain-matched and click-free. LISTEN auditions Band, Voice or Removed, latching on click and momentary on hold. LEARN plays the source, places the band where its esses actually live and calibrates the ride, and it refuses to guess when there is not enough sibilance to trust. The drawer holds the detector: Relative or Absolute threshold, a plus or minus 12 dB offset, Precise latency at exactly 191 samples at 48 kHz or Live at zero, detection in Stereo, Mid Focus or Sides, channel link, sense, and individual attack and release trims. There are 24 factory starting points named for sources, from close pop leads to hi-hats and steel-string guitar, each measured on its matched real take before it shipped. A and B hold two full snapshots with the session. The worst style costs 0.43 percent of one core at 48 kHz. In Live mode Surgical falls back to Split, because there is no lookahead, and the switch says so honestly. LEARN behaves the same way: given fewer than about 40 confidently sibilant frames it declines the material and changes nothing, rather than guessing at a calibration. Built and used at a working recording studio.