How to clean a noisy interview recording — ai denoise, no upload
- Step 1Drop the interview file — Open ai-noise-reducer and add the recording (MP3, WAV, M4A, FLAC, or a video interview). It stays on your device.
- Step 2Pick the output format — Set Output format: WAV (default) if you'll edit or transcribe next, or MP3/M4A for a finished file. There is no noise-strength control — RNNoise's suppression is fixed.
- Step 3Mind the stereo-to-mono sum — If you recorded interviewer and guest on separate channels, RNNoise will sum them to mono. To keep them separate, split first with channel-splitter and denoise each.
- Step 4Run the denoise — FFmpeg resamples to 48 kHz mono float; RNNoise removes the steady location bed on your CPU; the result is a 16-bit WAV (or 192 kbps MP3/M4A).
- Step 5Check both voices survived — Confirm the quieter speaker (often the interviewer, off-mic) is still clear and natural. The location hum should be gone.
- Step 6Tighten and level for publication — Cut dead air with silence-stripper, even out who-was-louder with speech-leveler, then hit a target with loudness-normalizer.
Interview noise sources vs RNNoise
RNNoise targets the steady bed. Other voices and one-off sounds are different problems.
| What's on the tape | RNNoise result | Handle elsewhere with |
|---|---|---|
| Cafe AC / espresso-machine hum | Strongly suppressed | — |
| Street traffic / wind rumble | Well suppressed under speech | High-pass via voice-eq for wind |
| Guest's room fan over the call | Strongly suppressed | — |
| A third person talking nearby | Preserved — it's speech to the model | No tool isolates speakers; edit manually |
| Chair scrapes, cutlery clinks | Largely remains (transient) | audio-splitter to cut |
| Quiet interviewer off-mic | Preserved, but may sound thin | speech-leveler to raise |
Interview cleanup chain
Recommended order for a publication-ready interview. Denoise first so later stages measure a clean floor.
| Step | Tool | Purpose |
|---|---|---|
| 1 | ai-noise-reducer | Remove the steady location noise bed |
| 2 | silence-stripper | Cut dead air and long pauses (-40 dB, 0.5 s) |
| 3 | speech-leveler | Even out loud/quiet speakers |
| 4 | loudness-normalizer | Hit -16 LUFS for publication |
| all-in-one | podcast-master | Run denoise -> strip -> normalise -> limit at once |
Cookbook
Field- and remote-interview scenarios. Each shows the noise problem, the format choice, and the next tool in the chain when denoise alone isn't enough.
Cafe interview with AC and machine hum
A one-mic recorder caught both voices plus the cafe's air-con and espresso machine. RNNoise removes the steady bed; the conversation stays clear.
Input: cafe-interview.wav (mono, 38 min, AC + machine hum) Output format: WAV RNNoise: suppress steady cafe bed Output: cafe-interview-clean.wav (mono, 48 kHz) Next: silence-stripper -> loudness-normalizer
Remote interview: guest's fan down the line
Recorded from a call where the guest's room fan bled into their mic. Denoise the whole mixed file, then level since the guest is louder.
Input: remote.m4a (mixed, guest louder, fan noise) Output format: WAV Step 1 ai-noise-reducer -> remote-clean.wav (fan gone) Step 2 speech-leveler -> even out guest vs host
Two-mic interview recorded in stereo
Host on left, guest on right. RNNoise sums to mono. If you want per-speaker editing, split first.
Input: two-mic.wav (L = host, R = guest) If mono is fine: denoise directly -> summed mono clean file To keep speakers separate: 1. /audio-tools/channel-splitter -> host.wav, guest.wav 2. Denoise each independently 3. Recombine in your editor
Prep a noisy interview for transcription
You need an accurate transcript. Removing the steady bed first lifts ASR accuracy. WAV out feeds most transcription engines.
Input: doorstep.m4a (traffic + wind) Output format: WAV RNNoise: strip traffic rumble Output: doorstep-clean.wav -> send to transcription Note: a cleaner floor reduces transcription errors.
Background chatter you cannot remove
A third person is audible behind your guest. RNNoise treats that as speech and keeps it — denoise won't isolate your two speakers.
Input: panel-corner.wav (guest + background chatter) RNNoise output: steady bed gone, BUT background voices remain (They are speech, not noise, to the model.) No tool here isolates one speaker; cut affected sections manually with /audio-tools/audio-splitter.
Edge cases and what actually happens
Two-mic stereo interview is summed to mono
By designRNNoise is mono-only, so a stereo interview (host left, guest right) is downmixed to one channel. If you need to edit speakers independently, split with channel-splitter first, denoise each, then recombine.
Background voices are not removed
Not supportedRNNoise removes steady non-speech noise. A third person talking is speech to the model and is preserved — there is no speaker isolation. Cut affected sections manually with audio-splitter.
Quiet interviewer sounds thin after denoising
ExpectedAn off-mic interviewer had a low signal-to-noise ratio, so removing the noise can leave their voice thin. Raise and even it with speech-leveler, and add warmth with voice-eq.
Chair scrapes and cutlery clinks remain
ExpectedThese are transients, not the steady bed RNNoise targets, so they largely survive. Edit them out with audio-splitter or audio-trimmer.
Wind rumble persists on an outdoor interview
ExpectedLow-frequency wind rumble sits below the speech band the model weights most, so it is reduced but not gone. Apply a high-pass with voice-eq to clean the low end.
Interview file exceeds the tier size cap
413-style rejectA long interview can exceed the per-tier file cap (Free 10 MB, Pro 50 MB). Export a smaller MP3, split it with audio-splitter, or upgrade to Pro+Media for files up to 100 GB.
Daily denoise run already used
Quota exhaustedFree allows one denoise per day, Pro five (reset at UTC midnight). For multiple interview parts in one day, merge them with audio-merger and clean in one run, or upgrade.
RNNoise WASM module fails to load
Error'RNNoise WASM module is unavailable.' means the ~85 KB module could not load — an old browser or a blocked dynamic import. Use a current browser and retry.
Output is 48 kHz mono, not the original layout
By designInput is resampled to 48 kHz mono for RNNoise, so the output is 48 kHz mono. Convert the rate afterward with sample-rate-converter if your delivery spec differs.
Frequently asked questions
Will it clean both the interviewer and the guest?
Yes — RNNoise removes the steady noise bed around all the speech, so both voices come out cleaner. It does not, however, separate the two speakers; if they're on one mixed track they stay mixed, and a quiet off-mic interviewer may need leveling afterward with speech-leveler.
Does cleaning the audio help transcription?
Yes, measurably. Automatic speech recognition struggles with a high noise floor. Removing the steady cafe/AC/traffic bed first usually reduces transcription errors, so export WAV and send the cleaned file to your transcription tool.
My interview was recorded in stereo with each person on a side — what happens?
RNNoise downmixes to mono, summing both channels. If a mono interview is fine, denoise directly. To keep speakers separate for editing, split with channel-splitter first, denoise each channel, and recombine.
Can it remove a third person talking in the background?
No. Background voices are speech to the model, not noise, so they're preserved. There is no speaker-isolation feature. Cut the affected sections manually with audio-splitter.
Is my interview uploaded anywhere?
No. RNNoise and FFmpeg 8.1 run as WebAssembly on your CPU; the recording is processed entirely in your browser. That keeps an off-the-record or embargoed interview private.
Can I adjust the amount of noise removal?
No. RNNoise is a fixed, trained network with no strength slider — the only control is output format. It applies the same suppression each time, which keeps a multi-part series consistent.
What order should I process an interview in?
Denoise first, then strip silence, then level the speakers, then normalise loudness. Denoising first means the silence detector and loudness measurement see a clean floor. You can run the whole sequence at once with podcast-master.
Wind noise is still there after denoising — why?
Low-frequency wind rumble sits below the speech band RNNoise weights most, so it's reduced but can remain. Apply a high-pass filter with voice-eq to clear the low end on outdoor interviews.
What formats can I import and export?
Import MP3, WAV, M4A, FLAC, or a video file (audio extracted). Export WAV (default), MP3, FLAC, or M4A. WAV is best if you'll transcribe or run further processing; lossy exports are 192 kbps.
How long an interview can I clean?
The file-size cap is per tier: Free 10 MB, Pro 50 MB, Pro+Media up to 100 GB. A long interview usually needs Pro+Media or an MP3 export / split first. Free is also limited to one run per day.
Can I clean the audio from a video interview?
Yes. Drop the video; FFmpeg extracts the audio, RNNoise denoises it, and you get a clean audio file. Re-attaching it to the video happens in your editor.
Why is the output 48 kHz mono?
RNNoise requires 48 kHz mono, so input is resampled and downmixed, and the output matches. Convert the sample rate afterward with sample-rate-converter if your delivery format needs 44.1 kHz.
Privacy first
Every JAD Audio tool runs entirely in your browser via FFmpeg (WebAssembly) and RNNoise. Your audio files never leave your device — verified by zero outbound network requests during processing.