How to remove fan and ac noise from recordings — free, local ai
- Step 1Drop the noisy recording — Open ai-noise-reducer and add a single file with the fan/AC noise (MP3, WAV, M4A, FLAC, or video). On the free tier the file must be 10 MB or smaller.
- Step 2Pick the output format — Choose Output format: WAV (default), MP3, FLAC, or M4A. There is no fan-specific setting and no strength slider — RNNoise's suppression of steady noise is fixed.
- Step 3Run your free daily denoise — The free tier allows one run per day. FFmpeg resamples to 48 kHz mono float and RNNoise removes the steady fan/AC bed on your CPU. Make it count — pick the take you most need cleaned.
- Step 4Expect mono output — RNNoise downmixes to mono. For a single voice over a fan, that is exactly what you want.
- Step 5Download and verify the bed is gone — Listen for the continuous hum to drop out. Transients (a door, a click) may remain — RNNoise targets steady noise, not one-off sounds.
- Step 6If 10 MB is too small, plan around the cap — Trim the recording with audio-trimmer, export a smaller MP3 first, or upgrade to Pro (50 MB) / Pro+Media (100 GB) for longer files.
Free vs paid AI-denoise allowance
AI denoise uses the aiDenoise preview quota. Daily counts reset at UTC midnight. These are the real numbers from the tier table.
| Tier | Runs per day | Max file size | Notes |
|---|---|---|---|
| Free | 1 | 10 MB | Enough to rescue a short take; resets at UTC midnight |
| Pro (£7/mo) | 5 | 50 MB | Five runs a day, larger files |
| Pro+Media (£19/mo) | Unlimited | 100 GB | No daily cap; long files and batches up to 100 files |
| Developer | Unlimited | 100 GB | Unlimited runs and batch |
Steady mechanical noise vs RNNoise
Continuous, broadband noise is the model's strength. One-off sounds are not.
| Noise source | Character | RNNoise result |
|---|---|---|
| PC / desk fan | Steady broadband hiss + whine | Strongly suppressed |
| Air conditioning / HVAC | Steady low rumble + airflow | Strongly suppressed |
| Coil whine / electrical hum | Steady tonal-ish hum | Well suppressed under speech |
| Server-room / appliance hum | Steady broadband | Strongly suppressed |
| Door slam, single click | Transient | Largely remains — not a steady bed |
| Sub-bass rumble below 80 Hz | Steady low rumble | Reduced; add a high-pass via voice-eq if needed |
Cookbook
Fan- and AC-noise removal, scoped to the free daily allowance where it matters. Each example notes the format choice and any cap you'd hit.
Short voice take next to a loud PC fan (free tier)
A 2-minute clip recorded beside a gaming PC under load. Under 10 MB as an M4A, so it fits the free daily denoise.
Input: take.m4a (mono, 2 min, ~3 MB) — fits free 10 MB cap Output format: MP3 RNNoise: remove steady fan whine + rumble Output: take-clean.mp3 (192 kbps) Free allowance used: 1/1 today
AC-unit hum on a webcam clip, cleaned from the video
A short talking-head clip with an air-con unit overhead. Drop the video — the audio is extracted and denoised.
Input: clip.mp4 (video, AC hum) Output format: WAV FFmpeg: extract audio -> 48 kHz mono float RNNoise: strip AC bed Output: clip-clean.wav (mono) Re-attach to the video in your editor.
Long recording too big for the free 10 MB cap
A 35-minute WAV is far over 10 MB. Two paths: shrink it, or upgrade.
Input: session.wav (45 MB) — exceeds free 10 MB cap Option A (stay free-ish): 1. Export a 192k MP3 elsewhere to shrink it, then 2. Denoise the smaller MP3 (still subject to 10 MB) Option B: upgrade to Pro (50 MB) or Pro+Media (100 GB)
Fan noise plus sub-bass rumble
A desk fan adds whine; the desk itself transmits low rumble into the mic. RNNoise handles the whine; a high-pass cleans the rumble.
Step 1 ai-noise-reducer fan.wav -> clean.wav (fan bed removed) Step 2 voice-eq clean.wav -> final.wav (high-pass < 80 Hz) RNNoise is strong on broadband; EQ finishes the sub-bass.
Two free runs needed in one day — plan it
You have two clips to clean but only one free run per day. Combine them first so a single run covers both.
Problem: 2 clips, free tier = 1 run/day Fix: 1. /audio-tools/audio-merger clipA + clipB -> combined.wav 2. Denoise combined.wav (one run) 3. /audio-tools/audio-splitter split back if needed Keep the merged file under 10 MB for the free cap.
Edge cases and what actually happens
Free file is over 10 MB
413-style rejectThe free aiDenoise preview caps the input at 10 MB. A longer or lossless recording exceeds it. Shrink to a smaller MP3 first, trim with audio-trimmer, or upgrade to Pro (50 MB) or Pro+Media (100 GB).
Free daily run already used
Quota exhaustedFree is one denoise per day, resetting at UTC midnight. Once spent, the run is blocked with an upgrade prompt. Merge multiple clips into one file with audio-merger to clean them in a single run, or upgrade for more runs.
One-off sounds (door, click) survive the fan removal
ExpectedRNNoise targets the steady fan/AC bed, not transients. A door slam or single click is one-off and largely remains. Cut those manually with audio-splitter or audio-trimmer.
Sub-80 Hz desk rumble is only partly reduced
ExpectedVery low rumble is steady but sits below the speech band the model weights most. It is reduced but may persist. Add a high-pass filter with voice-eq to finish the low end.
Stereo recording returns as mono
By designRNNoise is mono-only, so stereo is downmixed before denoising. For a single voice over a fan this is the right result. Split with channel-splitter first if you need stereo preserved.
RNNoise WASM 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. Update the browser or change networks and retry.
Output is 48 kHz mono regardless of source rate
By designInput is resampled to 48 kHz mono for the network, so the WAV output is 48 kHz. Convert to 44.1 kHz afterward with sample-rate-converter if needed.
MP3/M4A export bitrate is fixed at 192 kbps
By designThere is no bitrate control. Lossy targets are re-encoded at 192 kbps. For a different bitrate, export WAV and run bitrate-changer.
Frequently asked questions
Is it really free?
Yes, for short takes. The free tier gives you one AI-denoise run per day on a file up to 10 MB, resetting at UTC midnight. That covers a couple of minutes of audio without a card. Pro (£7/mo) raises it to five runs a day and 50 MB; Pro+Media (£19/mo) is unlimited up to 100 GB.
Why is fan and AC noise easier to remove than other noise?
Fans, AC, and HVAC produce steady, broadband, non-speech noise — a continuous bed with no melody. That is exactly what RNNoise was trained to separate from speech, so it suppresses it cleanly while leaving the voice. Transient or musical noise is much harder.
Do I have to upload the recording?
No. RNNoise and FFmpeg 8.1 run as WebAssembly on your CPU. The recording is processed entirely in your browser and never uploaded.
Can I set how strongly it removes the fan?
No. RNNoise is a fixed neural network — there is no strength slider. The only setting is output format. It applies the same trained suppression every time, which keeps results consistent.
My recording is bigger than 10 MB — what now?
Either shrink it (export a smaller MP3 or trim it with audio-trimmer) to fit the free 10 MB cap, or upgrade: Pro allows 50 MB, Pro+Media allows files up to 100 GB.
I have two clips but only one free run today — any workaround?
Yes. Merge the clips into one file with audio-merger (keeping it under 10 MB), denoise in a single run, then split them back with audio-splitter if needed.
Will it remove a one-off door slam or click?
Not reliably. RNNoise targets the steady fan/AC bed, not transients. A single slam or click is largely preserved — cut those manually with audio-splitter or audio-trimmer.
There's still low rumble after denoising — why?
Sub-80 Hz rumble sits below the speech band the model weights most heavily, so it is reduced but can persist. Apply a high-pass filter with voice-eq to clean the very low end.
Why is the cleaned file mono?
RNNoise processes a single mono channel, so stereo is downmixed first. For a voice over a fan this is the intended result. Use channel-splitter to keep stereo if you need it.
What output formats can I choose?
WAV (default, lossless 16-bit), MP3, FLAC, or M4A (AAC). Lossy formats are encoded at 192 kbps. WAV is best if you'll EQ or normalise next.
Does the free run expire or roll over?
It resets daily at UTC midnight and does not roll over — an unused run does not bank for tomorrow. Plan to use it on the take you most need cleaned.
What if the denoiser won't start at all?
If you see 'RNNoise WASM module is unavailable,' the WebAssembly module could not load — usually an outdated browser or a network blocking the dynamic import. Update to a current browser and retry.
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.