Description: "CleanTake removes steady room noise from voice recordings on your own computer — no upload, no account. Real before/after test (25 s) and the exact method."
tags: audio, opensource, ai, showdev
canonical_url: https://cleantake.bronzebasehq.workers.dev/blog/free-local-audio-denoiser/
Originally published on my blog.
CleanTake removes steady room noise — fan, air-conditioner hum — from recorded voice files, entirely on your computer. No upload, no account, no subscription. On a real 25-second recording tested September 28, 2026, it cut environmental noise by about 32 dB while the voice stayed intact. It cannot separate overlapping voices, and since October 2, 2026, there's a free public download — a portable ZIP or one-click installer for Windows.
Download CleanTake v1.0.1 — free
Does it actually work? Listen to the same 25 seconds, before and after
One take, one voice, recorded in a room with a fan running. The before file is the raw recording. The after file is the same take cleaned by the CleanTake pipeline. Listen to the steady hum underneath the voice in the first clip, and how much of it survives in the second.
🎧 Before — raw recording (MP3)
🎧 After — cleaned (MP3)
Noise type: fan / air-con hum · Test 2026-09-28, re-checked 2026-09-30 · Engine DeepFilterNet 0.5.6 · ffmpeg 8.1.2 · 25.0 s each · Linux CPU (public build targets Windows)
What to listen for: in the before clip, the hum sits behind the voice at a near-constant level — it never pauses. In the after clip, the voice keeps its tone and pacing, but the hum drops out of the pauses between sentences. Nothing was re-recorded, re-acted, or sweetened.
What kind of noise does it remove — and what can't it do?
It handles:
Steady room noise: fans, air-conditioners, electrical hum
Hiss from cheap microphones and interfaces
Voice recordings where the speaker is the only voice
It can't do (yet):
Overlapping voices. In a café-style test with four people talking over each other, it reduced everything evenly — voice included. It does not separate one voice from a crowd.
Real-time calls. It processes recorded files, not live audio. It is not a Krisp replacement for Zoom.
Music. It is tuned for speech; singing and instruments are not its job.
The honest version is the short one: CleanTake is a cleaner for environmental noise, not a separator for human noise. If your problem is a noisy room and a single speaker, that's exactly the case it passed. Everything else is on the limits page, including the failed café test.
Why clean audio locally instead of using a free online tool?
Free options exist, but they each cost something that isn't money. Checked September 30, 2026:
CleanTake
MyEdit (online)
sibilsoren/audio-denoiser (GitHub)
Where processing happens
Your machine
Your browser (per their FAQ)
Your machine
Account to download the result
No
Yes — 1 free download per day
No
Setup
Installer / portable ZIP
None
Python 3.11 + FFmpeg, terminal commands
Measured before/after proof on the page
Yes — above
Sample player, no method published
Before/after player inside the app
The pattern: cloud tools gate downloads behind sign-ins and daily limits; open-source repos work but demand a terminal. CleanTake's bet is the gap in the middle — proof-first page, one-click installer, your files never leaving the room.
How was this tested? Exact method, versions, and numbers
No cherry-picking claims without a method. Here is everything needed to reproduce the result:
Source: a real 25.0-second voice recording in a room with fan/air-con hum (MP3, 400 KB, 48 kHz mono). Not synthesized, not re-recorded.
Pipeline: ffmpeg → DeepFilterNet 0.5.6 → loudness normalize to −16 LUFS (I=−16, TP=−1.5, LRA=11). Output: MP3, 601 KB.
Result: environmental noise reduced by ≈32 dB, voice preserved — measured on this sample, 2026-09-28. Your mileage varies with the room.
Speed: the full pipeline processed this 25-second clip in about 5 seconds on a standard CPU with no GPU (timed 2026-09-30, Linux test machine).
Re-check 2026-09-30: an independent measurement of the noise floor in two speech pauses showed the floor dropping from −33.0 dB to −51.1/−59.3 dB — an 18–26 dB reduction under a different, stricter metric. Both are published so you can see the method matters.
Voice preservation: a listening assessment (the voice sounds the same), not a lab metric — labeled as such.
Who is this for — and who should skip it?
Podcasters and voiceover folks recording at home with untreated rooms — fan and AC noise is the single most common complaint.
Anyone shipping voice notes, course audio, or narration who wants a free local option without uploading client audio to a server.
Skip it if your audio has multiple people talking over each other, or you need real-time noise removal on live calls.
Get CleanTake
CleanTake v1.0.1 is out — free download for Windows, as a portable ZIP (no install) or a setup installer. 100% local: no account, nothing uploaded anywhere.
Download v1.0.1 — free · How to use it · DeepFilterNet on GitHub
Competing products named above were checked 2026-09-30 and described as found. CleanTake v1.0.1 is released for Windows; macOS/Linux builds are planned.
Top comments (4)
the method section is what makes this believable. exact versions, re-checked dates, and a listening assessment labeled as one instead of a lab metric - most launch posts give you a demo and ask for trust. the 'cleaner, not a separator' line is the sharpest sentence here, it tells me exactly when not to use it.
Thanks, hat line cost me the most rewrites, honestly. Every denoiser demo shows you the best 10 seconds
I figured the fastest way to earn trust was to say exactly where it fails. Crowded cafe? — wrong tool.
If you run it on your own recordings, I'm curious what breaks first.
crowded cafe should be a benchmark every denoiser ships with. my ceiling fan is already volunteering to find out what breaks first.
Haha — tell your ceiling fan it's about to become famous.
And honestly, "crowded café benchmark" is a great line. Might steal it for the README — with credit.
If the fan test produces anything weird (robot voice, underwater artifacts), I want to hear about it. Failure samples teach me more than clean ones.