Voice enhancer

Drop a voice memo, an interview or a podcast take and get a clean WAV back. Denoise takes the hiss and the room out and leaves your voice exactly as it was. Enhance goes further and rebuilds the speech through a neural vocoder — your recording, as if it had been made in a good room.

  1. Your recording

    Drop a voice recording here, or choose a file

    MP3WAVM4AOGGFLACMP4

    Your files are never uploaded — everything runs on your device.

  2. Choose the job

  3. Your file

Two different jobs, not two strengths

Denoise is a filter. A 41 MB model predicts which parts of the sound are noise and takes them away; every word that comes back is the word you said, in the voice you said it in. It runs faster than real time on an ordinary processor, so it works on any machine, including a phone.

Enhance is a rebuild. Five models — a denoiser, an encoder, a conditional-flow-matching network, a decoder and a UnivNet vocoder — resynthesise the speech from scratch, so what you download is a new waveform rather than a cleaned copy of the old one. That is why it can fix things a filter cannot: a muffled laptop mic, a hard bathroom echo, a phone call recorded through a table. It is also why the tool shows the original beside it every time.

The long version, with the five models in order and the arithmetic behind the numbers on this page: what an AI voice enhancer actually does — why a filter cannot fix a muffled mic, and where a rebuild lies.

Already have a clean recording and only want the hiss gone? The noise remover runs DeepFilterNet 3 on audio and video and is the lighter tool for that job. Want the words rather than the audio? Transcribe it.

The honest limits

The models are Resemble AI’s resemble-enhance (MIT), exported to ONNX and served from our own storage. Enhance is generative: it reconstructs your speech rather than recording it, so the result is not evidence of what was said. Sibilance, breaths and very quiet speech are where a rebuild is most likely to differ from the original, and long silences can pick up artefacts that were never in the room. When the recording itself matters, use Denoise.

Enhance is also graphics-card only, and not by preference: the four generative models measure 6.3 seconds of computing per second of audio on a processor, which is a tool nobody can use, so they are refused before the download instead of offered as a slow path. It needs roughly 6 GB of free graphics memory at once — an 8 GB card manages it with nothing else heavy running — and about 1.3 GB of one-time download. Denoise asks for 41 MB and no graphics card at all.

Why this, on-device

A voice recording is a person talking: an interview under embargo, a therapy note, a medical dictation, a family message from someone who is gone. Those are exactly the files people hesitate to hand to a website that will not say what happens to them. Your files are never uploaded — everything runs on your device. Every one of the five models runs in this browser tab, on your own graphics card or processor, and the recording never has to leave the computer it is already on.

Questions people ask

How is this different from the background noise remover?

YourDevice has two tools for a noisy voice recording, and they do opposite things to the same problem. /remove-background-noise/ runs DeepFilterNet 3, which filters noise OUT: what is left is your recording, minus the hiss and the hum, and every word is the word you actually said. This tool’s Enhance mode does not filter at all — it listens to the recording and re-renders the speech through a neural vocoder, so what comes back is a new waveform that sounds like you in a better room. Denoise here is the filtering job (a different model from DeepFilterNet, and a smaller download); Enhance is the rebuild. If the recording is basically fine and you just want it clean, the noise remover or Denoise is the right tool. If it was recorded in a bathroom on a laptop mic, Enhance is the one that helps.

What model is this?

The YourDevice voice enhancer runs Resemble AI’s resemble-enhance (MIT), exported to ONNX and served from our own storage. Denoise is its 41 MB denoiser UNet, which predicts a mask over the recording’s spectrum. Enhance adds four more graphs — an IRMAE encoder, a conditional-flow-matching network, an IRMAE decoder and a UnivNet vocoder — for about 1.3 GB in total. The last of those is what makes Enhance generative: UnivNet renders audio from scratch, it does not process yours.

Why does Enhance need a graphics card when Denoise does not?

Because Enhance and Denoise are different amounts of work, by two orders of magnitude. The denoiser is one pass over a ten-second window and runs faster than real time on a plain processor. Enhance runs sixty-four solver steps on every two and a half seconds of audio, plus an encoder, a decoder and a vocoder — measured on a processor, that is 6.3 seconds of computing for every second of audio, which is not a slower tool, it is an unusable one. So it is offered only where there is a graphics card to run it, and refused before the download rather than after it.

How long does it take, and how much does it download?

It depends which of the voice enhancer’s two modes you pick. Denoise is 41 MB and runs in a fraction of the recording’s length. Enhance is about 1.3 GB across five models and runs at roughly the length of the recording itself — measured at 0.77x real time on an RTX 2080, so a two-minute memo took about a minute and a half. It also needs roughly 6 GB of free graphics memory at once, which an 8 GB card has only if nothing else heavy is using it. Both downloads are one-time: the models stay on this device, so the second recording costs nothing to start.

Will it change what I said, or make me sound like someone else?

The voice enhancer’s Enhance mode keeps your words and your voice — that is what resemble-enhance was trained to do — but it is a reconstruction, not a recording, so it is not evidence. Sibilance, breaths and the very quietest speech are where a rebuild is most likely to differ audibly from what you actually said. For anything where the recording itself matters — an interview you will quote, a legal or medical note, a piece of proof — use Denoise, which never touches the voice, or the noise remover. The tool always shows both files side by side so you can hear the difference before you keep one.

Does anything get uploaded?

No. Your files are never uploaded — everything runs on your device. The voice enhancer’s models are downloaded once to this device and then run in this browser tab for every recording after that, including the one you just cleaned up. There is no account, no queue and no per-minute charge, because there is no server doing the work.