Mix Tools

GLASSCUT

A de-esser built on spectral level restoration rather than high-frequency compression.

£29 £39Launch price

See the measurements

Formats: VST3 · AU · Standalone
Operating system: macOS · Universal
Version: v0.1.0

What it does

The lisp is in the gain stage, not the detector.

A de-esser is not a high-frequency compressor with a good detector. It is a spectral level-restoration problem, and the artefacts that make de-essers sound bad come overwhelmingly from the gain application, not the detection. A perfect detector driving a fixed 6 kHz shelf still produces a lisp.

GLASSCUT shapes the attenuation to match the sibilant's own measured spectral envelope — a bell whose centre, bandwidth and depth track the measured contour, realised with time-domain filters. Because the notch matches the noise's own shape, the /s/ stays an /s/ and simply gets quieter. That is why it holds together at 12 dB where a shelf lisps at 6.

Transparency is the highest priority, so the architecture spends its budget on the gain stage at least as heavily as on the detector.

Key features

Harmonicity measured in the time domain

The hardest discrimination is /s/ versus a bright vowel — both put strong energy in 4–10 kHz, and level, centroid and band ratios all fail on a soprano through a bright condenser. The physics separates them cleanly: a vowel's HF energy is a glottal pulse train, periodic at F0; an /s/ is turbulent noise, aperiodic.

Forty times cheaper than the obvious approach

Resolving harmonic structure in the FFT would need a 43 ms window, and a short /s/ lasts 40 ms — unusable. Instead the HF envelope is band-passed, rectified, low-passed and decimated by sixteen, giving a ~3 kHz stream on which a 43 ms correlation window is 128 samples. Same measurement, computed from a 10.7 ms slice.

Spectral Contour Matching instead of a shelf

A fixed shelf or crossover removes the vowel's air and the cymbals along with the /s/. FFT bin masking smears in time and goes watery under heavy reduction. SCM shapes the reduction to the measured contour and avoids both.

Dual-rate detection

A fast path — time-domain filterbank, about 1 ms — gives a provisional gain. A slow path — STFT, about 5 ms — arrives to confirm or veto with confidence and frequency. Classification is inherently latent, because at the onset of an /s/ the analysis window is still half full of the preceding vowel. No algorithm beats that.

Breath is a separate class

Not a sub-case of sibilance, and acted on only when you ask. A breath is aperiodic broadband noise, so the periodicity test cannot separate it from an /s/ — only level, tilt and centroid can.