Filery

Audio Noise Reduction Beta

Reduce background noise in recordings. Works best on consistent, steady noise.

Worth knowing before you startNoise reduction works well on consistent background noise — room tone, fan hum, microphone self-noise, and low-level hiss. It is less effective on intermittent sounds (traffic, keyboard clicks, a dog barking), music in the background, or noise that is loud relative to the signal. Aggressive settings will introduce artefacts: a bubbling or watery character on speech, and loss of natural ambience. There is no setting that removes all noise without affecting the signal.
Choose audio filess or drop them anywhere on this page MP3 · WAV · OGG · OGA · M4A · AAC · FLAC · OPUS · WMA · AIFF · +4 Processed on your device — never uploaded

How to audio noise reduction

  1. Add a recording with unwanted background noise.
  2. Try strength 5 first. If it is not enough, increase it — but listen for the characteristic watery artefacts that appear at high settings.
  3. Download and check in headphones before deciding on the final setting.

About this tool

Background noise in recordings comes from the environment: air conditioning systems, computer fans, the natural ambience of a room, and the self-noise of the microphone itself. These noises are characteristically steady and spectrally consistent, which is exactly what noise reduction algorithms exploit. A spectral gate measures the noise floor during quiet sections, builds a model of the noise spectrum, and attenuates those frequencies across the whole file.

The honest limit of this approach is that "noise" and "signal" are not always well-separated in the frequency domain. A recording made in a reverberant room contains reflections of the speaker's voice in every direction — some of which look like noise to a spectral analyser. Reducing the noise also slightly reduces those reflections, which changes the acoustic character of the recording. At moderate settings this is usually acceptable. At aggressive settings it becomes the dominant character of the output — that watery, processed sound that reveals the algorithm.

Questions

Can it remove music playing in the background?
No, not effectively. Spectral gating distinguishes between consistent noise (same frequencies at relatively constant level) and signal (variable, transient, spectrally rich). Background music has all the characteristics of a signal — it varies over time and covers a wide frequency range. Removing it would require separating two mixed signals, which is the audio-vocal-remover problem and requires a stem separation model.
Why do I hear a "watery" sound at high settings?
At aggressive reduction levels the spectral gate removes frequency bins that it identifies as noise even when they contain residual signal. The result is an incomplete spectrum — certain frequencies disappear in and out as the gate opens and closes, which produces a modulated, warbling artefact colloquially called "musical noise". It is the signature of over-processing. Reduce the strength to the lowest setting that achieves acceptable noise reduction.
What does the RNNoise model do differently?
RNNoise is a recurrent neural network trained specifically to distinguish speech from non-speech in real-time phone calls. It applies frame-level suppression based on the learned characteristics of voice versus noise rather than a fixed spectral gate. For voice recordings in a noisy environment it typically produces fewer artefacts than aggressive spectral gating at the same noise reduction level. It only works for speech — it has no model for music or other sound types.