How to denoise and normalize your podcast in one step — free
- Step 1Drop the raw recording — Drag the unprocessed recording onto the tool — audio or video. This is the recording straight out of your recorder or DAW edit, before any cleanup.
- Step 2Pick the loudness target — Choose the platform target the normalise stage should hit: Apple Podcasts (-16 LUFS), Spotify / YouTube (-14 LUFS), Amazon Music (-14 LUFS / -2 dBTP) or EBU R128 (-23 LUFS).
- Step 3Pick the output container — MP3, WAV, FLAC or M4A. For maximum quality through a host's re-encoder, choose WAV or FLAC; for direct publishing, MP3 or M4A.
- Step 4Run the combined pass — Click Master. RNNoise denoises (resampling to 48 kHz mono), the silence stripper trims gaps, then the 2-pass loudnorm sets your target and ceiling — all without stopping for input.
- Step 5Read the loudness report — Check the measured integrated LUFS, LRA and true peak. They confirm the normalise stage landed on target. A 'dynamic' normalization type means the source range was too wide for a clean linear pass.
- Step 6Download the clean, leveled file — Download the denoised and normalised mono file. If denoise alone wasn't enough for a very harsh recording, run the result through standalone tools for finer control (see edge cases).
Why this order: denoise then normalize
The chain runs denoise -> silence -> normalize. Here's what each step contributes and why running them out of order produces worse results.
| Step | What it does | Why the order matters |
|---|---|---|
| 1. RNNoise denoise | Removes steady background noise from the speech | Run first: a clean noise floor lets the loudness meter measure the voice, not the hum |
| 2. Silence strip | Removes gaps below -40 dB lasting 0.5 s+ | Run mid-chain: stops long quiet gaps from skewing the integrated loudness measurement |
| 3. EBU R128 normalize | Sets integrated LUFS + -1 dBTP ceiling | Run last: measures the already-clean audio so the target is hit accurately |
What RNNoise fixes — and what it doesn't
RNNoise is a speech-trained neural network. It is strong on steady noise and weak on transients/reverb. Set expectations accordingly.
| Noise type | RNNoise result | Better tool / fix |
|---|---|---|
| Fan / AC / PC hum | Removed well | This chain handles it |
| Constant hiss / mic self-noise | Removed well | This chain handles it |
| Room reverb / echo | Largely remains | Acoustic treatment at source |
| Plosives / pops | Not its job | High-pass / edit in DAW |
| Harsh sibilance (s/sh) | Not its job | de-esser |
| Mouth clicks | Mostly remains | Manual edit / silence-stripper for gaps |
Loudness targets the normalize stage can hit
Presets read from the engine. The normalize stage uses the I/TP/LRA from whichever you pick.
| Preset | Integrated | True peak | LRA |
|---|---|---|---|
| Apple Podcasts | -16 LUFS | -1 dBTP | 11 LU |
| Spotify / YouTube | -14 LUFS | -1 dBTP | 11 LU |
| Amazon Music | -14 LUFS | -2 dBTP | 11 LU |
| EBU R128 broadcast | -23 LUFS | -1 dBTP | 7 LU |
Cookbook
Denoise-and-normalize runs from real noisy recordings. Each shows the noise problem, the target, and the cleaned-and-leveled result.
Bedroom recording with AC hum
The classic home-studio problem: a steady air-conditioner hum under an otherwise good vocal take, plus an inconsistent level.
Input : home-ep.wav (constant ~120 Hz AC hum, level drifts) Target : Apple Podcasts (-16 LUFS) Stage1 : RNNoise removes the AC hum continuously (no gate pumping) Stage2 : silence strip trims long pauses Stage3 : loudnorm -> -16.0 LUFS / -1.1 dBTP Result : home-ep-normalized.mp3 (mono, clean, on-target)
Noisy laptop-mic interview
An interview recorded on a built-in laptop mic — hiss, fan noise, and a quiet guest. Denoise plus normalize fixes both the floor and the level.
Input : interview.m4a (hiss + fan, guest 6 dB quieter than host)
Target : Spotify / YouTube (-14 LUFS)
Result : interview-normalized.m4a (mono, -14.0 LUFS)
Note : Loudness now matches across speakers because the integrated
measure averages the whole episode. For per-speaker evenness,
add speech-leveler before this step.When denoise alone over-processes the voice
On a very noisy source, RNNoise can introduce a slight 'watery' artefact. If that happens, denoise and normalize separately so you can A/B the denoise strength region by region.
Symptom: voice sounds slightly watery after the one-click run Reason : source noise floor was extreme; RNNoise worked hard Do instead: 1. ai-noise-reducer (denoise alone, audition it) 2. loudness-normalizer (set -16 LUFS on the cleaned file) This lets you keep or discard the denoise before committing to a level.
Reverb that denoise won't remove
A bathroom-echo recording. RNNoise removes the hiss but the echo remains — it's not a dereverb.
Input : echoey-room.wav
After chain: hiss gone, BUT room echo still audible (expected)
Reality: RNNoise targets steady noise, not reflections.
Fix : record in a treated space; or accept the echo. The normalize
stage still lands the level correctly on the de-hissed audio.Loudnorm goes dynamic on a wide-range source
A recording that swings from whispers to laughter. After denoise, the loudness range is still wider than 11 LU, so normalize falls back to dynamic mode.
Report : Normalization Type: DYNAMIC, source LRA 15.2 LU Effect : final LUFS may drift slightly from -16.0 Fix for a clean linear result: 1. denoise + level first (ai-noise-reducer -> speech-leveler) 2. then loudness-normalizer for the LUFS target The combined chain still produces a usable file; this just tightens it.
Edge cases and what actually happens
Output is mono after denoise+normalize
By designThe RNNoise denoise stage resamples to 48 kHz mono, so the combined output is mono. For voice podcasts this is correct. If you specifically need stereo denoise+normalise, there's no stereo path in this chain — denoise is mono-only here, so normalise a stereo file separately with loudness-normalizer and accept no AI denoise on it.
RNNoise leaves room reverb in place
Limited effectRNNoise is trained to remove steady noise, not reflections. Echo and reverb largely survive the denoise stage. The normalisation still works on the de-hissed audio, but the room sound remains. Treat the room acoustically or re-record for a genuinely dry result.
Denoise introduces a 'watery' artefact
Limited effectOn extremely noisy sources RNNoise can over-suppress and produce a faint watery quality on speech. Because the chain runs denoise automatically, you can't dial it back here. Use the standalone ai-noise-reducer, audition it, and only then normalise — giving you a chance to reject the denoise.
Loudness measured wrong because of a noisy floor — fixed by ordering
SupportedThis is exactly the problem the chain solves: if you normalise BEFORE denoising, the hum inflates the measured loudness and the voice ends up too quiet. Running denoise first (as this chain does) means the loudness pass measures clean speech. So this 'edge case' is handled by design.
Loudnorm switches to dynamic mode
ExpectedIf the denoised source still has a loudness range wider than the preset's target (11 LU, or 7 LU for EBU), FFmpeg uses dynamic normalisation and the final LUFS can drift. The report flags it. For a clean linear pass, level the dynamics first with speech-leveler.
Sibilance is worse after denoising
ExpectedRemoving background noise can make harsh 's' sounds more prominent because they're no longer masked. RNNoise doesn't de-ess. Run the master result through the de-esser if sibilance stands out after cleanup.
Free preview file too large for denoise
Tier limitThe free daily preview caps input at 10 MB (and the audio family at 50 MB / 30 min). A typical 30-minute WAV exceeds 10 MB. Compress to MP3 first with bitrate-changer, or upgrade to Pro (200 MB / 120 min) / Pro+Media (100 GB, unlimited).
Music in the recording sounds damaged
Limited effectRNNoise treats anything non-speech as noise, so background music or musical stings can be degraded. If your episode has music beds, denoise the voice-only segments separately or skip denoise and just normalise the full mix with loudness-normalizer.
Want denoise but no loudness change
Out of scopeThis tool always denoises AND normalises. If you only want denoising (keeping your existing level), use the standalone ai-noise-reducer instead — it skips the loudness pass entirely.
Frequently asked questions
Why does denoise have to come before normalize?
Background noise inflates a loudness meter — a hum or hiss adds energy the meter counts as 'loudness'. If you normalise a noisy file, the meter targets the noise+voice combined, leaving the voice too quiet once the noise is later removed. Cleaning first means the loudness pass measures the actual speech, so the target is accurate. This chain enforces that order.
Is RNNoise a noise gate?
No. A gate cuts everything below a threshold, which causes pumping and chops off quiet speech. RNNoise is a neural network trained on speech that continuously separates voice from steady noise, so it removes hum and hiss even underneath talking, without the on/off artefacts of a gate.
What noise can it actually remove?
Steady, broadband background noise: fans, air conditioning, computer hum, mic self-noise/hiss. It is not designed to remove echo/reverb, plosive pops, mouth clicks, or sibilance — those need acoustic treatment, editing, or the de-esser.
Why is the result mono?
The denoise stage resamples to 48 kHz mono before running RNNoise, so the combined output is mono. That's ideal for spoken-word. There's no stereo path through this chain; for stereo you'd normalise separately with loudness-normalizer and skip the AI denoise.
How accurate is the loudness?
It uses 2-pass EBU R128: pass one measures the real integrated LUFS, true peak and range; pass two applies a corrective gain plus the -1 dBTP ceiling. Unless the source range is too wide (which triggers dynamic mode), the output lands within a fraction of a LU of the chosen target.
Can I just denoise without changing the volume?
Not with this combined tool — it always normalises too. For denoise-only, use the standalone ai-noise-reducer, which keeps your existing level.
Will it remove the hum but leave echo?
Yes — that's expected. RNNoise removes steady noise like hum and hiss but is not a dereverb, so room echo will remain. The normalisation still works correctly on the de-hissed audio.
Does any of my audio get uploaded?
No. RNNoise and FFmpeg 8.1 both run in your browser via WebAssembly. The recording is processed entirely on your device, which matters for noisy interviews containing private or unreleased material.
My voice sounds watery after denoising — why?
That happens when the source is very noisy and RNNoise has to suppress aggressively. Because this tool denoises automatically, switch to the standalone ai-noise-reducer so you can audition the denoise before committing, then normalise the version you like.
Will denoising make sibilance worse?
It can — removing masking noise exposes harsh 's' sounds that were previously hidden. RNNoise doesn't de-ess. Follow up with the de-esser if sibilance becomes noticeable.
What's the largest file I can denoise and normalize?
Free preview: 10 MB. Pro: 200 MB and up to 120 minutes. Pro+Media and Developer: up to 100 GB streamed with no duration cap. The size and duration limits are checked separately, so a long but small MP3 may pass where a short WAV won't.
Should I add a compressor too?
If your speech swings widely between quiet and loud, add speech-leveler before this step. It evens out the dynamics so the normalise pass can run in clean linear mode and hit the target precisely.
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.