How to Remove Background Noise from Voice Recordings
The fan was humming, the traffic was rumbling, and your voice still came through fine. Here's how neural denoising separates the two — and how to do it without handing your recording to a stranger's server.
What's Actually in That Recording
Every recording is a mixture: your voice plus a background of steady noise (fans, HVAC, room tone, traffic rumble) and transient noise (keyboard clacks, paper, door sounds). Traditional noise gates chop quiet passages but leave noise under your voice — the hard part is cleaning the moments when you're actually talking.
That's exactly the problem neural denoisers were built for: separate the speech from the noise in every single moment, not just the silences.
RNNoise: A Neural Network That Listens
RNNoise is an open-source noise suppressor from Xiph.Org — the foundation behind the Opus codec that powers Zoom, Discord, and Spotify's audio. Inside is a recurrent neural network trained on thousands of hours of speech and noise, learning to recognize what a voice sounds like even under heavy interference.
The process, every 10 milliseconds:
- The frame is analyzed for speech-like features (pitch, harmonics, spectral shape).
- The network estimates how much of the frame is voice versus noise.
- The noise component is suppressed; the speech is passed through.
It runs in real time on a phone, which means it's trivially fast in your browser's WebAssembly. Our test: 11 dB of pink noise removed from a speech recording, speech correlation preserved at 0.83.
What It Fixes — and What It Can't
| Noise type | RNNoise? | Notes |
|---|---|---|
| Fans, HVAC, room hum | ✓ Excellent | Steady broadband noise — the ideal case |
| Traffic, outdoor rumble | ✓ Good | Suppressed while speech stays clear |
| Keyboard & mouse clicks | ✓ Good | Transients are reduced noticeably |
| Second person talking | △ Limited | It's speech — the model keeps it |
| Loud music | ✗ Damages | Speech model treats music as noise |
| Clipping / distortion | ✗ Can't fix | Already baked into the waveform |
Rule of thumb: RNNoise is for voices. Podcasts, voiceovers, calls, video dialogue — yes. Music — no.
Best Practices: Clean at the Source First
- Mic closer, room quieter. Six inches from a good mic beats a foot away in a loud room.
- Turn off the noise sources. The AC, the fridge, the fan, the traffic-facing window — cheaper than any plugin.
- Record at a healthy level — peaks around −6 to −3 dB, never clipping.
- Denoise, then gate. Neural denoise for the moments you're speaking, then a light gate for the silences.
- High-pass at ~80–100 Hz afterward to remove rumble the denoiser left behind.
For heavily noisy material, running the denoiser twice is safe — the second pass rarely hurts speech and can shave off another few dB.
Do It in Your Browser — Nothing Uploaded
Your recording may be a confidential client call, an unreleased episode, or a demo you don't want in anyone's training data. The Prescosoft Voice Denoiser keeps it all local:
- RNNoise compiled to WebAssembly — the neural network runs in your browser tab.
- Record from your mic or drop in a file — MP3, WAV, M4A, OGG, FLAC.
- A/B compare, watch the waveform clean up, and download a 48 kHz WAV.
Close the tab and the recording is gone. No account, no watermark, no limits.
Frequently Asked Questions
What is the best way to remove background noise from audio?
For voice recordings, a neural denoiser like RNNoise is the best first pass — it removes 10-15 dB of steady noise (fans, hum, traffic) while preserving speech. Record as cleanly as you can first (close mic, quiet room), then denoise. For extreme cases, pair a denoiser with a noise gate and a high-pass filter around 80-100 Hz to remove rumble.
Is RNNoise good?
Yes — RNNoise is the open-source neural noise suppressor from Xiph.Org, the foundation behind the Opus codec used by Zoom, Discord, and Spotify. It runs a recurrent neural network in real time and is widely considered the best free speech denoiser. In our verification test it removed 11 dB of pink noise from a speech recording while keeping speech correlation at 0.83.
Why does my audio need to be mono for denoising?
RNNoise's neural network is a single-channel speech model — it analyzes one stream of samples. Stereo recordings are downmixed to mono for processing (most podcast and voiceover work is mono anyway), then exported as a clean mono WAV. If you need stereo output, denoise the mono mix and re-sync it with the original in your editor.
Can noise removal fix badly clipped audio?
No. Clipping (distortion from recording too hot) is already baked into the waveform — a denoiser separates noise from speech but cannot reconstruct clipped peaks. If your recording clips, lower the input gain and re-record, or use a declipper/declicker before denoising.
Is it safe to upload recordings to online noise removal sites?
Uploading an unreleased recording to a web service hands your audio — possibly confidential client calls or unreleased content — to a third party. Some free services use uploaded audio for training data. A local denoiser like the Prescosoft Voice Denoiser runs entirely in your browser: the recording never leaves your device, no account, no tracking.
How much noise can RNNoise remove?
Typically 10-15 dB of steady background noise — enough to make a fan- or hum-filled recording sound clean. It handles fans, traffic, room tone, and keyboard sounds well. It won't remove loud music, a second person talking, or heavily distorted audio. For music, use a music-specific tool — RNNoise is trained on speech.
Silence the noise, keep your voice
RNNoise neural denoising, entirely in your browser. Nothing uploaded, ever.