How to find out whether the hiss was in your recording or added by the edit

Last updated 2026-09-17 · 8 min read

What is the short version?

A hiss you can hear in the delivered file has two possible origins, and they need opposite fixes. Denoising a recording that was already clean damages the speech for nothing; rebuilding the chain when the microphone really was noisy fixes nothing at all.

1. Give Vizard Agent the raw audio and the export.
2. Ask it to measure the noise floor in the pauses of both.
3. Decide what to do once you know which it is.

What do you need before you start?

Both ends of the chain. Vizard Agent compares the recording against the delivered file, so the raw microphone files matter more in this job than in almost any other audio work — without them there is simply nothing to compare the export against.

• The raw microphone files. Each one separately.
• The delivered export. The version you can hear it in.
• Where you hear it. A timecode helps.
• How it sounds. Hiss, hum, crackle, whine.
• What processing was applied. If you know.

What do you type into Vizard Agent?

Ask for the diagnosis before the cure. "Clean up this hiss" gets you a denoised file whether or not denoising was the right answer, and a denoised file always sounds slightly worse than the original it came from.

Prompt

Variants worth knowing:

• "Before cleaning anything." Puts the diagnosis first.
• "In the silent passages." Where a noise floor is measurable.
• "Whether we made it." Names the hypothesis worth testing.

What does Vizard Agent actually do?

Here is the Vizard Agent sequence on a real episode with a suspected electrical noise in it. Almost every step is measurement, and the actual cleaning amounts to a single decision taken right at the end with a good deal of evidence behind it.

1. Locates silent passages in the raw material.
2. Measures the noise floor in those silences on each microphone.
3. Measures the same thing on the edited episode.
4. Views the spectrogram of a silent passage on both microphones.
5. Scans both microphones across the episode for bursts of static.
6. Produces a spectrogram of the suspect passage in the export.
7. Compares the spectra of speech, music and raw microphone at that point.
8. Checks how the export's speech track was actually built.
9. Rebuilds the speech track from the raw microphones and compares.
10. Tries several cleaners, listens to each, and picks the most natural.

Step three is the comparison the whole job turns on. A noise floor is a number, so "the raw sits at one level and the export sits ten decibels above it in the same silence" is a finding rather than an impression.

Step eight is where manufactured noise usually comes from. A track assembled with the wrong gain staging, an aggressive processor or a bleeding second microphone will add a floor that was never recorded, and the fix is to rebuild rather than to denoise.

Step ten is worth insisting on even after a clean diagnosis. Denoisers differ, and the difference between them is audible on speech, so the choice should be made by listening to several rather than by accepting the first.

What does the result look like?

An answer and then a repair. You are told where the noise came from, what the two noise floors measured, and which of several cleaning approaches sounded most natural — rather than being handed a quieter file and left to wonder what it cost.

When does this not work well?

Some recordings have no usable silence in them at all. Continuous music, a live room, or a speaker who never pauses for breath leaves Vizard Agent nowhere to measure a floor, and the comparison that the whole method depends on becomes much harder to make.

• No pauses. Nothing to measure the floor in.
• Music throughout. It masks the floor at both ends.
• Raw files unavailable. No comparison is possible.
• Already-processed "raw". It is not actually the recording.
• Intermittent bursts. A floor measurement will not find them.

How do you fix a result that came back wrong?

Say whether the noise is still there or whether the cure has hurt the voice. Those two reports lead in opposite directions, and Vizard Agent still has the measurements it took, so it can move either way without going back to guesswork.

• "I can still hear it." A different cleaner, or the chain is rebuilt.
• "The voice sounds underwater." The cleaning was too strong; it is backed off.
• "It comes and goes." Bursts rather than a floor; the scan is re-run.
• "It is on one microphone only." That track is rebuilt or excluded.

How does Vizard Agent compare to doing it yourself?

By hand you reach for a denoiser, because that is what the symptom suggests. It works, the hiss goes, the speech goes slightly dull, and nobody ever finds out that the recording was clean and the mix had been adding the noise all along.

Common questions

Why do I need the raw files? Because the comparison is the diagnosis. Without them Vizard Agent can only clean.

What is a noise floor? The level of the quiet parts. Vizard Agent measures it only where nobody is speaking.

Can processing really add hiss? Yes. Vizard Agent checks gain staging, aggressive processors and microphone bleed.

What if there is no silence? Harder. Vizard Agent finds the shortest pauses it can and says when there are too few.

Does denoising always hurt the voice? A little, always. Vizard Agent weighs that cost, which is why the order of work matters.

Can it try several cleaners? Yes, and it should. Ask Vizard Agent to let you hear each one.

What about a hum rather than a hiss? A different shape on the spectrogram, often a cleaner fix, and Vizard Agent uses the same method.

Can it find intermittent bursts? Yes. Vizard Agent scans the whole track rather than measuring a single floor.

What if only one microphone is bad? Vizard Agent measures them separately and can rebuild from the clean one.

Will it tell me the numbers? Yes. The two noise floors Vizard Agent measured are the point of the exercise.

Can it rebuild the mix from raw? Yes, if that turns out to be the right answer.

Does this apply to an old archive recording? Yes, though there may be no raw to compare against.

How long does the diagnosis take? Minutes. Far less than re-delivering a video that still has it.

Should I always check before cleaning? When you have the raw files, yes. It is the cheapest thing in the job.