Spectral repair has transformed dialogue editing since iZotope RX introduced it to professional workflows. What previously required ADR — a line ruined by a phone ring, a truck passing, a chair squeak — can now often be fixed in minutes at the edit workstation. But spectral repair tools are among the easiest to misuse in audio post, and over-processed dialogue is almost as unacceptable as the original problem it was supposed to fix.

Understanding the Spectrogram

Spectral repair works on the frequency-time representation of audio — the spectrogram. Every sound has a specific spectral signature: dialogue appears as formant bands (the characteristic horizontal structures of voiced speech), while broadband noise appears as a diffuse texture, and tonal interference (like a camera motor hum) appears as a horizontal line at a specific frequency. Learning to read spectrograms fluently is the most important skill in spectral repair work.

In iZotope RX, the Spectral Repair module works by identifying a region containing the problem sound and replacing it with content interpolated from the surrounding audio. For short, contained problems (a single click, a brief tonal intrusion) this works extremely well. For sustained problems (persistent traffic noise, continuous hum) it requires a different approach — Spectral De-noise, not Spectral Repair.

When to Use Spectral Repair vs De-noise

Spectral Repair is ideal for: impulsive events (clicks, pops, crackles), brief tonal intrusions (phone rings, beeps, equipment noise lasting less than half a second), and specific frequency events that can be isolated visually on the spectrogram. De-noise is ideal for: continuous broadband noise (air conditioning, location hum, camera noise), persistent tonal noise (power line hum at 50/60 Hz and harmonics), and sustained interference that Spectral Repair cannot interpolate around without audible artefacts.

The Over-Processing Problem

The most common mistake with spectral repair tools is applying too much processing in pursuit of a perfectly clean signal. A voice that has been aggressively de-noised loses its natural texture — consonants become smeared, sibilants develop artefacts, and the fundamental character of the voice changes in a way that's immediately perceptible to anyone listening critically. This is called "processing artefacts" and it's a QC failure as surely as the original noise problem.

The professional principle is to ask: does this line need to be completely clean, or does it just need to be mixable? A line with moderate background noise that can be managed with careful fader automation during the mix does not need aggressive spectral treatment. Save the heavy processing for lines where the noise level genuinely prevents the line from working in context.

Dialogue Isolate: Possibilities and Limits

iZotope RX's Dialogue Isolate uses machine learning to separate dialogue from background noise by training on the spectral characteristics of the human voice. It's remarkable when it works and clearly audible when it fails. The tool works best on voice-only or voice-dominant material with moderate background noise. It struggles with: highly reverberant dialogue (the room tone shares spectral characteristics with the voice), overlapping voices, very loud background noise, and non-speech vocal sounds (singing, whispering, laughter).