You just pulled a promising EVP session off your recorder, but underneath the possible voice sits a wall of hiss, HVAC rumble, and room noise. That’s exactly the problem izotope rx voice denoise was built to solve. This tool isolates unwanted background sound and strips it away without shredding the fragile whisper you’re actually trying to hear, which matters a lot more in EVP work than in typical podcast cleanup.
In this guide, you’ll learn the exact workflow for applying Voice De-noise to raw field audio: how to select a clean noise profile, set reduction amounts that preserve faint anomalies instead of erasing them, and avoid the over-processing that turns a real voice into digital mush. We’ll walk through settings specific to EVP audio, since aggressive denoising can destroy the very evidence you’re hunting for.
Whether you’re reviewing a session with an entry-level recorder or running gear from our Haunt Gears shop, the cleanup process is the same. We’ll cover practical steps you can apply tonight, plus a few mistakes that cost investigators clear evidence. By the end, you’ll know how to turn noisy field recordings into audio you can actually analyze and share.
Why EVP recordings need careful noise cleanup
Field recordings fight you from the start. Unlike a podcast recorded in a treated room with a shock-mounted mic, your EVP sessions happen in basements, cemeteries, and drafty old houses where every surface adds its own sound signature. HVAC systems cycle on mid-session, floorboards creak, and your own recorder’s preamp hisses at gain levels most studio engineers would never touch. That layered noise floor is why generic "remove background noise" presets fail so often on paranormal audio. They’re tuned for human speech patterns recorded close-mic, not the faint, irregular anomalies you’re chasing.
Gear quality plays into this too. A budget digital recorder pushed to high gain to catch a whisper across a room will pick up more self-noise than a purpose-built EVP recorder designed for low-level capture. If you’re still shopping, check out our EVP recorder reviews to see which models keep noise floors lower before you even open RX. Cleaner source audio always beats trying to fix a bad recording after the fact.
The over-processing trap
Here’s where most investigators go wrong: they crank noise reduction to maximum because the recording sounds cleaner that way. It does sound cleaner, but that’s the problem. Aggressive denoising doesn’t just remove hiss, it removes low-amplitude detail, and a possible EVP is almost always low-amplitude. Push the reduction too far and you’ll get a suspiciously smooth, artifact-free track with nothing usable left in it. You traded evidence for silence.
Aggressive noise reduction doesn’t clean EVP audio, it deletes it.
This is the core reason Voice De-noise in iZotope RX works better than blunt instrument tools. It’s built around learning the specific noise profile of your recording environment first, then subtracting only that profile, rather than applying a one-size-fits-all filter across the whole frequency range. Done carefully, it separates the room tone from the anomaly instead of flattening both into the same processed mush.
Common noise sources you’ll actually encounter
Knowing what you’re dealing with before you open RX saves time. Different noise types respond to different settings, and mixing up your approach wastes a session.
| Noise source | Typical characteristic | Best RX approach |
|---|---|---|
| HVAC/ventilation hum | Steady, low-frequency rumble | Voice De-noise with a long Learn sample |
| Electrical hum (60Hz) | Consistent tonal buzz | De-hum module, layered after De-noise |
| Recorder self-noise | Broadband hiss, consistent level | Voice De-noise, moderate reduction |
| Footsteps/clothing rustle | Short, irregular transient spikes | De-click or manual spectral repair |
| Wind or fabric brushing mic | Low-frequency gusts, inconsistent | Spectral repair, avoid heavy De-noise |
Why this matters for credibility
Beyond just hearing the voice clearly, clean audio matters when you share findings. A recording riddled with hiss invites skepticism before anyone even considers whether the voice is genuine. A properly de-noised clip, processed with restraint and documented settings, holds up better to scrutiny from other investigators and skeptics alike. That’s the real payoff of learning this workflow properly instead of mashing a preset button and hoping for the best.
Step 1. Import and prep your EVP audio in RX
Start by opening iZotope RX and dragging your raw session file straight into the editor. Skip any pre-processing you might normally run in your DAW first. Trimming, normalizing, or applying EQ before RX sees the file can smear the noise floor and make Learn mode less accurate later. Work from the original recorder file whenever possible, even if it’s a clunky WAV straight off an SD card.
Once the file loads, switch to the Spectrogram view instead of the default waveform display. This is non-negotiable for EVP work. A waveform shows you volume over time, but a spectrogram shows you frequency content, which is where you’ll actually spot anomalies buried under noise. Faint vocal-range activity often shows up as a visible smudge or wisp in the spectrogram long before you can consciously pick it out by ear.
Set your zoom and resolution correctly
EVP anomalies are frequently short, sometimes under half a second, so your default zoom level will hide them. Zoom in horizontally until individual seconds stretch across a good portion of your screen. Then bump up the FFT size in the spectrogram settings (found in the display preferences) to improve frequency resolution. A higher FFT size trades some time resolution for clearer frequency detail, which helps you separate a faint voice from a similarly pitched noise like wind or hum.
Scrub for a clean noise sample before touching any tools
Before you apply anything, listen through the full session once, start to finish, without skipping. Mark timestamps mentally or with RX’s marker tool where you hear:
- A stretch of at least 2-3 seconds with no talking, footsteps, or obvious movement
- Consistent background character (the same hum or hiss level throughout)
- No overlapping voice, even faint, at the very start or end of the stretch
That clean stretch is gold. It’s what you’ll use to teach RX what "noise" sounds like in Step 2, and a sloppy selection here undermines every setting you dial in afterward.
The quality of your noise sample determines the quality of every result that follows.
Save a duplicate of your raw import before doing anything destructive. RX lets you work non-destructively in most modules, but a backup copy protects you from an accidental render that overwrites your only recording of a real anomaly.
Step 2. Capture a noise profile with Learn mode
With your clean stretch identified, select it precisely in the spectrogram or waveform view. Click and drag across just that noise-only region, nothing more. Including even a half-second of voice or footstep noise in your selection teaches RX the wrong lesson, and it’ll try to remove parts of your actual anomaly later thinking it’s part of the room tone.
Open the Voice De-noise module from the RX process menu, then click Learn. This is where the magic happens: RX analyzes the frequency content of your selected noise sample and builds a profile of exactly what the background sound looks like across the spectrum. It’s not guessing based on general presets, it’s measuring your specific room, your specific recorder, your specific HVAC hum on that specific night.
A precise noise sample teaches RX exactly what to remove, and nothing else.
Give Learn mode enough to work with
A single second rarely cuts it. Aim for 2-3 seconds minimum of pure noise, and longer if your session allows it. More data means RX builds a more accurate statistical model of the noise floor, which translates directly into cleaner separation once you apply the tool to the full recording. If your recording only offers a half-second gap between sounds, keep hunting through the session for a better stretch rather than settling.
Verify before moving forward
After clicking Learn, RX displays a noise profile graph showing amplitude across frequencies. Look for a profile that’s smooth and consistent rather than jagged with sharp spikes. Jagged spikes usually mean your selection caught a transient sound, like a distant clunk or a breath, that snuck into your "clean" sample.
If the graph looks off, don’t move forward yet. Go back and:
- Re-select a tighter or different stretch of noise
- Zoom in further to confirm no voice or movement bleeds into the selection
- Re-run Learn mode and compare the new profile against the first
Take this step seriously. A rushed or sloppy noise profile is the single most common reason investigators end up with either an over-processed track that erases the anomaly, or an under-processed one that still buries it in hiss. Everything downstream in Step 3 depends on getting this profile right, so don’t rush past it just to hear results faster.
Step 3. Apply Voice De-noise to isolate the voice
With your noise profile learned and verified, select the entire session, or at least the stretch you suspect holds the anomaly, and get ready to apply the process. Don’t touch the sliders yet. RX defaults to a moderate reduction setting out of the box, and that default is a reasonable starting point for a first pass on EVP audio. Resist the urge to crank anything up before you hear what a conservative pass actually does to the recording.
Click Render Preview rather than committing to a full render immediately. This lets you audition the processed audio against the original without permanently altering your file. Toggle between the processed and unprocessed versions using the preview controls, and listen specifically for whether the faint anomaly you flagged earlier still has body to it, or whether it’s thinned out and ghostly in a way that suggests over-processing.
Listen in short loops, not the whole file
Looping a two or three second window around your suspected anomaly beats playing through the whole session every time you tweak a setting. Long playback sessions fatigue your ears and make subtle differences harder to catch. Set a loop region right on the moment in question, then A/B between processed and raw audio repeatedly. Your ear gets sharper at spotting artifacts the more directly you can compare the same few seconds back to back.
If the anomaly sounds thinner after processing than before, you’ve already gone too far.
Watch the spectrogram while you listen
Don’t rely on your ears alone. Keep the spectrogram visible during preview playback and watch what happens to the frequency content where your anomaly sits. A clean Voice De-noise pass removes the diffuse haze around the voice while leaving the voice’s own frequency pattern intact and visible. If that pattern starts to fade or break apart on screen, the setting is too aggressive even if it sounds acceptable through speakers.
Run this preview-and-compare cycle at least twice before committing to render. First pass, confirm the noise floor actually dropped. Second pass, confirm the anomaly survived that drop. Only move to fine-tuning in Step 4 once you’ve got a baseline processed version that keeps both of those things true at the same time.
Step 4. Fine-tune threshold, reduction, and filters
Once your baseline preview holds up, it’s time to adjust the actual controls instead of relying on defaults. The Reduction slider in Voice De-noise controls how much of the learned noise profile gets subtracted, measured in decibels. Start around 6-9 dB of reduction for typical EVP sessions rather than the higher values you’d use on clean podcast audio. Push past 12-15 dB and you’re usually into territory where faint anomalies start losing their edges, even if the overall track sounds quieter and more polished.
Six to nine decibels of reduction preserves detail that twelve or more will erase.
Threshold works differently than reduction, and mixing up the two costs people results. Threshold sets the level below which RX treats sound as noise worth touching at all. Set it too low and you’ll process sounds that aren’t actually noise, including quiet portions of a real voice. Set it too high and RX barely touches the recording, leaving hiss intact. Nudge the threshold down in small increments, previewing after each change, until you find the point where the noise floor drops but your flagged anomaly’s waveform still shows movement in the spectrogram.
Use the frequency filters to protect vocal range
Voice De-noise includes filter controls that let you exclude certain frequency bands from processing entirely. Human speech, including whispered EVP-style anomalies, typically sits between roughly 300 Hz and 3,000 Hz. You can widen or narrow that protected band depending on what you’re hearing, but don’t skip this step. Setting a filter that shields the core vocal range means RX focuses its reduction on the frequencies outside that window, like low-end rumble or high-frequency hiss, without touching the range where a whispered voice actually lives.
Adjust in small steps, not big jumps
Restraint matters more here than anywhere else in the workflow. Move each slider in small increments, five to ten percent at a time, and preview after every change rather than making three adjustments at once and guessing which one helped.
- Increase reduction by small amounts until hiss noticeably drops
- Check the spectrogram for signs the anomaly is thinning
- Adjust threshold only after reduction feels close to right
- Widen vocal-range filters slightly if the voice sounds muffled
Stop adjusting the moment the anomaly sounds clearer without sounding artificial. That’s your target setting, not the point where the track sounds cleanest overall.
Step 5. Layer De-hum or De-click for stubborn noise
Voice De-noise handles broadband hiss well, but it wasn’t built to erase every noise type you’ll run into on a field recording. Electrical hum from bad wiring, old fluorescent fixtures, or a poorly shielded recorder cable sits at a fixed frequency, usually 60 Hz in US buildings, and it barely budges under a standard Voice De-noise pass. Sharp transients like a distant door latch, a knee pop, or static from clothing rubbing the mic also slip past it since they’re brief spikes rather than a steady noise floor. Once your voice track sounds clean on the broad strokes, check for these leftover problems before you call the session finished.
Spotting and removing electrical hum
Open the De-hum module and look at the spectrogram for a thin, perfectly horizontal line sitting at a consistent frequency, usually somewhere around 60 Hz with harmonics stacked above it at 120 Hz, 180 Hz, and so on. De-hum lets you target that exact frequency and its harmonics without touching anything else in the spectrum, which matters because a whispered anomaly can sit dangerously close to those same low frequencies. Set the fundamental frequency slider to match what you see on screen, enable harmonic removal, and preview before committing.
Hum removal only works when you target the exact frequency, not a broad range around it.
Cleaning up clicks without touching the voice
Run the De-click module next if you’re hearing sharp pops, crackles, or short mechanical noises layered on top of the voice track. De-click scans for sudden amplitude spikes and smooths them without affecting the surrounding audio, which makes it safer to apply broadly than Voice De-noise ever is.
Work through these modules in order for the cleanest result:
- Confirm Voice De-noise output still preserves the anomaly
- Apply De-hum if a steady tonal line shows up in the spectrogram
- Apply De-click for isolated pops or transient spikes
- Preview after each module, not just at the end
Skipping straight to a full render without this layered check is how investigators end up submitting a clip to a review thread, only to have someone point out the 60 Hz buzz they never noticed under headphones.
Step 6. Render, compare, and confirm your EVP catch
Once your modules stack up clean, it’s time to commit. Select the full session in RX and choose Render rather than Render Preview this time. Save the processed file under a new name instead of overwriting your original, something like session_2026-07-10_denoised.wav. You want both versions sitting on your drive side by side, because the raw file stays your evidence of record and the processed file becomes your listening copy.
Set up a proper A/B comparison
Load both files into RX or your DAW on separate tracks, lined up sample-for-sample so playback stays in sync. Loop the exact window where your suspected anomaly sits, then switch between tracks using a single mute toggle instead of scrubbing manually each time. This keeps the comparison fast and consistent, which matters because ear fatigue creeps in fast when you’re replaying the same three seconds for the twentieth time.
A real EVP survives the cleanup with its shape intact. A processing artifact usually doesn’t.
Checklist before you call it confirmed
Run through this list before you treat the clip as usable evidence:
- Does the anomaly still have a consistent waveform shape in both versions, not just a smoothed blob?
- Does the voice pattern in the spectrogram match between raw and processed, just with less surrounding haze?
- Does the sound survive at multiple playback volumes, not just when boosted loud?
- Can a second listener, ideally someone who hasn’t heard your theory about what it says, pick out the same sound?
- Does the timing of the anomaly line up with anything visible in your session notes, like a question you asked?
That last point matters more than people give it credit for. A processed clip that happens to line up with a question you asked out loud carries far more weight than a stray noise floating in silence.
Document your settings
Write down the reduction amount, threshold, and any filter ranges you used, along with which modules you ran and in what order. Keep this note attached to the file. If another investigator or a skeptic asks how you processed the audio, having exact settings on hand shows you followed a repeatable method rather than tweaking sliders until something sounded convincing. That documentation is often the difference between a clip people take seriously and one they dismiss outright.
Trusting what you hear after cleanup
Good cleanup work doesn’t manufacture evidence, it reveals what was already sitting under the noise floor. Voice De-noise gives you a repeatable way to strip hiss and hum without flattening the faint sounds you’re actually investigating, as long as you resist the temptation to over-process for a cleaner-sounding file. Learn mode, careful reduction settings, and layered De-hum or De-click passes matter more than any single slider position.
The habits from this guide, patient noise sampling, small adjustments, honest A/B comparisons, will do more for your credibility than any single processed clip ever could. Consistent methodology separates investigators whose findings hold up from those who get dismissed after one skeptical listen.
If your recordings keep fighting you before you even open RX, the problem might start at the mic. Browse the paranormal research devices in our shop built to capture cleaner audio from the start, so cleanup becomes a formality instead of a rescue mission.
