Studio Sound

The input is the voice track Video Cleanup produced: mono, 48 kHz, cut to the plan with 20 ms fades at every join, and otherwise exactly as recorded. A stereo input is folded to mono first. Nothing else is needed: no script and no finished render.

The track is sliced into 0.4-second windows. The noise floor is the RMS level of the 3rd-percentile window, the speech level is the 85th percentile, and the gap between them is the signal-to-noise ratio, the most useful single number in the pass. The same measurement also reports peak level, samples at full scale grouped into runs, DC offset, octave-band energy of noise and speech separately, narrow peaks at 50, 100, 150, 200, 60 and 120 Hz against their neighbours, spectral balance, sibilance, and the spread of 3-second levels across the whole lecture.

Full-scale samples are counted and grouped before any decision about clipping. Isolated runs of a few samples are the peak grazing full scale. They cannot be heard, and a de-clipper would replace real signal with interpolation for no benefit, so de-clipping is not run. The real peak problem is inter-sample overshoot, which the limiter and true-peak target handle at the end of the chain. Mains hum is notched only if a harmonic stands more than 9 dB above its neighbours. DC offset matters only above ±0.001, and the high-pass removes it anyway.

The 80 quietest 0.4-second windows, 32 seconds of material, become the noise profile. A generic preset would remove the wrong thing. Because cleanup has already trimmed long silences and cut the lead-in and tail, these windows come from real gaps between words in the lecture that will actually ship.

Spectral noise subtraction reduces the profiled noise by 75% of its amplitude, about 12 dB. The mask is smoothed across 400 Hz of frequency and 64 ms of time, in 2048-point frames with a 512-sample hop. Unsmoothed masks cause the watery, swirling artefacts on breaths and sibilants that students notice far more than a quiet room tone. A lecture that sounds slightly roomy is fine; one that sounds underwater is not.

A high-pass at 75 Hz removes desk rumble, footsteps and air handling. Three cuts take out the boxiness of a clip-on microphone: 2 dB at 180 Hz, 3.5 dB at 380 Hz and 1.5 dB at 900 Hz. A 3 dB lift at 3.2 kHz adds presence and intelligibility. A 4.5 dB high shelf from 7.5 kHz restores some of the air a lavalier under clothing loses, and a narrow 1.5 dB cut at 6.2 kHz keeps that lift from turning harsh. The shelf comes after the denoise, or it would simply amplify hiss.

A downward expander runs before any compression. It leaves everything from −24 dBFS upwards alone, pushes −34 dBFS down to −41 and −42 dBFS down to −62, with a 6 dB soft knee, a 20 ms attack and a 250 ms release so word tails are not chopped. Without it, the compressor's make-up gain lifts the residual noise between words straight back up and undoes the noise subtraction. On a measured lecture this one stage was worth 10 dB of final signal-to-noise.

Gentle compression at 2:1 above −20 dBFS, with a 12 ms attack, 220 ms release and 3.5 dB of make-up, evens out a leaned-in sentence against a trailing one. A limiter at about −1 dBFS (5 ms attack, 60 ms release) catches remaining transients. Loudness normalisation then sets the whole lecture to −16 LUFS integrated, a −1.5 dBTP true-peak ceiling and a 7 LU loudness-range target. −16 LUFS is the spoken-word and e-learning convention, and suits both a phone speaker and a lecture-hall PA.

The finished track is measured with the same windows as the source, then run through a full loudness meter for integrated loudness, loudness range and true peak. It is written as mono 48 kHz 16-bit audio. The job doctor then confirms it is the same length as the cut voice track to within 0.02 s. From then on, every effect and the finished film use this file and never the recording's own sound.