Noise removal
A neural de-noiser separates speech from everything behind it — traffic, air conditioning, computer fans, room reverb — instead of gating the quiet parts and leaving the loud ones untouched.
Nine tools, 0 bytes uploaded
Clean, cut and convert — without uploading a thing.
Strip out noise, breaths and dead air. Compress, convert or reframe video. It all happens on your own computer — drop a file in, tick what you want, and get a clean one back in minutes. No upload queue, no pay-per-hour billing, and no stranger's server in the middle of your work.
Download for Windows
Free — 2 hours a month, no account needed. v1.0.3. See pricing →
- No account needed to start
- Full audio cleanup, not a trimmed-down one
- Cancel any time, keep what you paid for
Audio 3
Video
1 file ready
Noise removal Breath removal Silence removal Audio enhance Loudness normalise Format convert
interview-take-3.wav
- Noise floor
- −31 dB −68 dB
- Breaths
- 48 found none left
- Dead air
- 2m 14s 11s
- Loudness
- drifting −16 LUFS
- voiceover-04.wavqueued
- client-cut-final.mp4video
Clean up everything
- 2 hrs free every month, at full quality
- 9 tools six for audio, three for video, one app
- Your CPU no queue, no GPU, no waiting your turn
- 0 bytes of your media sent to us, ever
- Windows 10 & 11
- Runs on CPU — no GPU required
- No account needed to start
Hear it
A real recording, before and after
The same recording, straight from the microphone and after one pass through the app. The pauses and dead air are cut out too, which is why the refined clip is shorter.
Original
0:00
Refined
0:00
A real recording, processed by the application. Nothing else was done to either clip for this page.
Before and after
Drag the handle. The gaps are what you stop cutting by hand
Left of the seam is the recording as it arrived: a noise floor under every word and two minutes of dead air spread through it. Right of the seam is the same file after one press — same peaks, same words, nothing between them.
Raw
Refined
- Noise floor
- −31 dB → −68 dB
- Breaths
- 48 found → none left
- Dead air
- 2m 14s → 11s
- Loudness
- drifting → −16 LUFS
Clean up everything
Five audio tools, one press, in the order that actually sounds right
Run separately and in the wrong order, these tools fight each other — a de-noiser chasing a breath the silence pass already cut, a loudness stage normalising to a floor that is about to be removed. So they do not run separately. One press runs the lot as a single job, the audio stays lossless the whole way through, and it is encoded once at the end.
-
Noise
Room tone, hiss, hum and fan noise pulled out from under the voice.
-
Enhance
The voice brought forward, so what is left is a recording rather than a survivor.
-
Breaths
Inhales found and removed — including the ones synthetic voices insert.
-
Silence
Dead air detected by speech, not by volume, and tightened to the pace you set.
-
Loudness
Normalised to a broadcast target so one episode does not open louder than the last.
Each step goes green as it finishes — and says so plainly when it had nothing to do.
What it does
One app for the audio and the video
The cloud cleanup tools do audio and stop. The free video tools do video and have no AI audio at all. A single delivery usually needs both, and needing two applications to make one file is the problem this was built to end.
Audio
Breath removal
Every inhale located and taken out at a strength you choose. The tool AI voice work needs most: synthetic narration inserts breaths that were never drawn, and nothing gives the source away faster.
Silence trimming
Gaps found by a speech detector rather than a volume threshold, so a quiet sentence is never mistaken for a pause. Set how tight you want the pacing and it holds it across the whole file.
Audio enhance
Body and clarity restored to a thin recording. The pass that turns a laptop microphone into something you can put in front of an audience.
Loudness normalise
Measured to the LUFS standard the platforms actually use, then corrected — so a series sounds like a series and nobody reaches for the volume between episodes.
Format conversion
WAV, MP3, FLAC, M4A, AAC, OGG and Opus, at the bit depth and sample rate your editing suite wants — converted once, at the end, from lossless.
Video
Compress
A four-gigabyte master cut down to something you can actually send, with the quality target in your hands rather than a preset's.
Convert
MP4, MOV, MKV and WebM, in the codecs the destination expects. One master file becomes every delivery version in a single queue.
Smart reframe
16:9 turned into 9:16 by following the face and cutting at the scene changes, not by dropping a fixed crop box over the middle and hoping.
How it works
Three steps, then it gets out of your way
No upload, no queue position, no waiting for a server in another country to finish with your recording.
-
1
Drop the files in
One file or forty. Everything you drop is included by default, and unticking one leaves it in the list without processing it this time — so a batch is edited, not rebuilt.
-
2
Tick what you want
Pick individual tools, or press Clean up everything and take the whole audio chain. Each tool's settings open underneath the tool itself, not in a dialog somewhere else.
-
3
Watch it work, locally
Steps go green as they finish, on your own CPU, at the speed of your own machine. Then the folder opens with the finished files in it.
Under the hood
Real models, running on your CPU
Nothing here is a filter chain wearing an AI label. These are the same published models the cloud services build on — they simply run on the processor you already own instead of one you rent by the hour.
- Full-band speech separation for noise removal — the voice lifted away from the room, not a gate muting the quiet parts.
- Speech detection for silence, so a pause is found by whether anyone is speaking rather than by how loud the moment is.
- A model trained on breaths for breath removal, rather than a filter tuned to a frequency band.
- Subject tracking and scene-change detection for reframing, so a 9:16 crop follows the person and cuts where the shot cuts.
- No GPL or AGPL code. FFmpeg ships as an unmodified LGPL shared library, which is safe for commercial work; everything else is MIT or Apache-2.0.
At a glance
- Platform
- Windows 10 & 11
- Where it runs
- your machine, CPU
- Files uploaded
- none
- Audio tools
- 6
- Video tools
- 3
- Batch processing
- yes
- Free allowance
- 2 hrs / month
- Runs on
- your own CPU
- Telemetry
- licence & usage time only — never your media
Why not the cloud
The difference running it yourself makes
Cloud cleanup tools charge by the hour because they are paying for the hour. Nobody is paying for yours.
| Cloud cleanup services | Sonic Refine | |
|---|---|---|
| Where your recording goes | Uploaded to someone else's server | Nowhere. It stays on your disk |
| Waiting | Upload, then a queue, then a download | It starts the moment you press it |
| What you pay for | Every hour you process, forever | A licence, plus a free monthly allowance |
| Video | Not offered — a second tool entirely | Compress, convert and reframe, same app |
| Client confidentiality | An NDA question you have to answer | Not a question. Nothing leaves the room |
| File size limits | Almost always | Your free disk space |
| If the service shuts down | You lose the tool | It keeps working. It is on your computer |
Who it is for
Built for people shipping on a deadline
AI voice creators
Synthetic narration arrives with inserted breaths, uneven pacing and drifting loudness. All three are one press here, and none of them cost an upload.
The tell that gives the source away, gonePodcasters
An hour of two-microphone conversation cleaned, tightened and normalised to a consistent target, without the hour being metered by anyone.
A weekly episode that costs nothing per hourVideo editors
One master turned into every delivery version — codec, size and aspect ratio — with the dialogue cleaned in the same pass rather than a separate round trip.
Two applications collapse into one queueAnyone under an NDA
Client footage that contractually cannot be uploaded to a third party is exactly the footage a local tool can process. Nothing to disclose, nothing to negotiate.
No third party to name in the agreement
Pricing
Pay for the software, not for the servers
There are no servers to pay for, which is the whole reason a licence here buys so much more processing than an hourly plan does. The free allowance runs the same models on the same machine — a licence buys hours, video up to 4K, every export format, and more files in one go.
Free
Two hours a month, no account needed.
$0.00 per month
- 2 hours of processing per month
- MP3 320 audio export
- 720p video, watermarked
- One file at a time
Basic
Twenty-three hours a month, clean 1080p.
$9.00 per month
- 23 hours of processing per month
- MP3 320 and WAV export
- 1080p video, no watermark
- 2 files per run
Pro
Seventy-five hours, 4K, and smart reframe.
$19.00 per month
- 75 hours of processing per month
- Every export format
- Up to 4K video
- Smart reframe
Studio
A hundred and fifty hours, on six machines.
$39.00 per month
- 150 hours of processing per month
- Every export format
- Up to 4K video
- Smart reframe
- Cancel any time — we never see your card
- Renewing extends your key, it does not replace it
- Expiry drops you to the free allowance — it never locks you out
Proof
The only review worth reading is your own file
You will not find a customer count or an award badge on this page, because we do not have either and a number nobody can check is worth less than none. What you can check is the software: install it, run the worst recording you own through it, and judge the output.
Download for Windows-
2 hours a month, free The full audio chain, the same models, no account to open.
-
No account to try it The free tier needs no key, so it never waits on our server at all.
-
Cancel whenever you want Card licences renew monthly until you stop them, in one click from your account, and you keep the month you have paid for. Bank transfer and wallet licences never renew at all.
Questions
The things people ask first
Do my files really never leave my computer?
They do not. Every pass — noise, breath, silence, enhance, loudness, and all of the video work — runs as a local process on your own machine. The app talks to our servers for two things only: checking your licence, and downloading the AI models once. Neither of those carries your media, and neither of them can, because the audio never enters the request.
What does the free allowance actually include?
2 hours of processing every month, running the same models a paid licence does, on your own machine, with no account and no nagging overlay. Audio comes out as MP3 320. Video is written at up to 720p and carries a small corner logo, and smart reframe is a Pro feature. A licence lifts all of that — but the cleanup itself is never the part being held back.
Why is it so much cheaper than the cloud tools?
Because the expensive part of a cloud audio service is the compute, and here it is your compute. A service that processes an hour of audio for you has to pay for that hour and charge you for it. We are selling you software, and the hardware it runs on is already sitting under your desk.
How fast is it?
It depends on your processor, which is the honest answer — a recent multi-core machine gets through a chain considerably faster than a five-year-old laptop. What it never does is queue: there is nobody ahead of you, and your upload speed is not part of the equation at all.
Does it need an internet connection?
It needs one to activate, to fetch the models the first time, and to keep your remaining processing time in step with your account. Your media never travels over it — every pass runs on your own machine.
Your remaining processing time is counted on our server rather than on your machine, which is why the app reports what it used. The free tier needs no key and no account at all.
Is this a subscription?
Only if you pay by card, and you can stop it in one click. A card licence renews each month until you cancel; cancelling stops the next charge and you keep the month you have already paid for. Paying by bank transfer or a mobile wallet never renews — there is no way for those to charge you. Either way, renewing extends the key you already hold — same key, same activated devices, nothing to paste in again.
Can I use it on client work?
Yes, and the local processing is usually the reason people can. There is no GPL or AGPL code and no GPL model anywhere in the build — FFmpeg ships as an unmodified LGPL shared library, and everything else is MIT or Apache-2.0 — and material that contractually cannot be uploaded to a third-party service never needs to be. You remain responsible for holding the rights to whatever you process.
Can I move it to a new computer?
Yes, and without asking us. Release the old device from your account, activate the new one, and carry on. Each plan includes a number of self-service moves per 30 days.
Clean a real file before you decide anything
Install it, drop in the worst recording you have, and press one button. Buy a licence only once it has earned one.