How to remove background noise from a voice recording on Android

What you need

A 64-bit Android phone running Android 8.0 or later, Noise Remover, and a recording with a voice in it: a voice note, a podcast take, a voice-over, a lecture or a video. No account and no connection are needed: the cleaning runs on the phone, and your audio is never uploaded.

The engine is RNNoise, an open-source neural network from Xiph.Org trained to tell speech from noise. It listens in 10 ms slices and adapts as it goes, so it needs no sample of the noise and keeps up with noise that changes, like traffic or typing. It is built for one voice at a time.

The free version gives 3 cleanings a day, each covering the first 3 minutes of the file, with no watermark; Pro removes the limits and the ads.

Several files can also be cleaned in one batch.

Step by step

1. Bring the recording in

   Tap Record to record in the app, Import audio to pick a file (MP3, M4A, AAC, WAV, FLAC, OGG, Opus, AMR, 3GP), or Import video to clean a video’s soundtrack. You can also share an audio or video file to Noise Remover from a messaging app, a voice recorder or a file manager. About to record? The recording tips matter more than any setting.

2. Let the cleaning run

   The cleaning starts on its own (by default at 80% intensity, with the ready-to-publish level and the low-cut filter on), and a waveform fills as it goes. On the free version, a file longer than 3 minutes is cleaned up to that point.

3. Compare before and after

   Play the result and switch between Original and Cleaned. The switch changes track at the same moment, with no gap, and both play at about the same loudness, so what changes is the noise, not the volume. Use headphones if you can, and listen to the pauses and the ends of words: that is where problems show first.

4. Set the intensity

   The slider runs from 0% (the original) to 100% (the full reduction), and you hear it change while the audio plays. 80% suits most recordings. If the voice sounds watery, lispy or clipped, go down to between 50 and 70%; for loud, steady hum or hiss, try 90 to 100%.

5. Check the two finishing switches

   Ready-to-publish level brings the loudness to −16 LUFS without letting the peaks pass −1 dBTP. Low-cut filter removes the rumble below 80 Hz, under the voice. Leave both on for most recordings; switching the low-cut runs the cleaning again, which takes a moment on a long file.

6. Trim the start and the end

   Drag the two handles on the waveform to cut the fumbling at the start and the silence at the end. The parts outside are dimmed, playback stays between the handles, and the export uses the same range.

7. Export, then share or save

   Share sends the cleaned file to any app; Save to device writes it to the folder you choose. By default you get an M4A file (AAC, 48 kHz, 128 kb/s) named after the original with “(clean)” added, and no watermark.

Besides M4A, the export can be WAV for further editing, MP3 for older players, or Opus for small files on Android 10 or later: tap the line above the export buttons to choose.

Each cleaning stays in History, on the phone, with a copy of its original: reopen it later to change the intensity or export it again.

Hear what a cleaning does

The sample below is a voice next to a fan, cleaned at the default 80%. Switch while it plays and listen to what stays: the breaths, the word endings, the tone of the voice.

What it removes, and what it can’t

In our lab test, RNNoise, the engine inside Noise Remover, removed 15 to 43 dB of steady and outdoor noise in the speaker’s pauses (hiss, hum, fans and air conditioning, rumble, traffic, wind, rain), while the voice stayed within about 1 dB of its level, 1.6 dB at worst. Keyboard clicks came down less, by 11 to 13 dB, with the voice untouched. Other voices were barely reduced, by 0 to 11 dB, and room echo stayed. How we measured.

Choosing the intensity

The intensity slider sets a cap on how far the noise can be turned down. The cap rises evenly from 0 dB at 0% towards 36 dB, and at 100% it is lifted. Noise Remover does this by mixing the original back in, far below the cleaned voice and delayed by the engine’s 20 ms so the two line up, which also means no part of the voice can be turned down by more than the cap.

The default, 80%, is a cap of about 29 dB. In our lab test, speech quality peaked with the reduction capped around 24 to 30 dB, while removing everything took out only a little more noise. Move the slider when you hear a reason to:

• 50 to 70% when the voice sounds watery, metallic or lispy, when breaths and word endings vanish, or when the recording was almost clean already.
• 80% for most voices recorded near a fan, air conditioning, traffic or wind.
• 90 to 100% for loud, steady hum or hiss, as long as the voice still sounds like you.

Normalization and the low-cut

Normalization makes the result loud enough to publish. Noise Remover measures the final mix (the cleaned voice at the intensity you chose, over the trimmed range) and applies one gain to reach −16 LUFS, unless the peaks would pass −1 dBTP first. There is no compressor or limiter, so the voice keeps its natural dynamics, and a recording with sharp peaks may land a little below −16 LUFS, which is normal. The preview already plays at the exported level. LUFS explained has the details.

The low-cut is a high-pass filter at 80 Hz, applied before the noise removal. It takes out the rumble of traffic, air conditioning, handling and wind, below the fundamental of most speaking voices. If a very deep voice sounds a little thinner, try it off; switching it runs the cleaning again from the original, which the app keeps for that purpose.

Cleaning the sound of a video

Import video reads the soundtrack of a video (MP4, 3GP, MKV, WebM, MOV) and cleans it like any recording; the picture is not touched. You export the cleaned sound as an audio file: in your video editor, place it under the clip and mute the clip’s original sound. More in how to remove noise from a video.

With Pro, you can also export the video itself with its cleaned soundtrack, as an MP4, and skip the editor.

Troubleshooting

The voice sounds robotic or watery

The reduction is cutting into the voice, usually because the noise was nearly as loud as the voice. Lower the intensity to between 50 and 70% and compare again: a little noise comes back, and much more of the voice does.

The result is too quiet

Check that Ready-to-publish level is on: it brings the export to −16 LUFS, or as close as the peaks allow. If the tip says your voice came out quieter than expected, something overlaps it (another voice, music), and a lower intensity helps.

There is still noise

Try a higher intensity. If the voice suffers before the noise is gone, the noise was as loud as the voice and no setting will separate them cleanly: re-record closer to the phone.

The voice sounds distant, as if in a bathroom

That is echo, the room’s reflections of your own voice, not background noise. A noise remover treats it as speech and leaves it at any intensity. See echo or background noise? for what works.

People talking or music are still there

Other voices are speech, so an engine built to keep speech keeps them, and when they overlap yours it can turn your voice down a little too. Songs with vocals behave the same way. Neither comes out cleanly: record away from conversations, and switch the music off.

If you can still re-record

Cleaning sounds most natural on a recording that was decent to begin with. If you can do the take again, four habits do more than any setting:

• Hold the phone about a hand span from your mouth, speaking toward its bottom edge.
• Switch off the fan, the air conditioning and anything else that hums, just for the take.
• Pick a furnished room (curtains, a sofa, a bed, a closet of clothes) rather than a bare one or a bathroom.
• Turn on airplane mode or Do Not Disturb, so no notification lands in the middle.

The other eight are in how to record clean audio with your phone.