How to remove background noise from a voice recording
Fan hum, street traffic, tape hiss — some noise can be reduced and some cannot. Knowing which is which saves you from ruining a good recording.
Know what you are dealing with
Noise reduction works very differently depending on the type of noise. Steady, predictable sounds — air conditioning, computer fans, electrical hum, tape hiss, wind — can be filtered effectively. The software learns the noise profile and attenuates it without touching the voice too much.
Unpredictable sounds are a different story. A dog barking, a door slamming, someone talking in the next room: these occupy the same frequency range as your subject and change over time. No simple filter can separate them. Being honest about this distinction is the single most useful thing to know before you start.
Start with the source
If you still have the original file, always work from that. Every round of processing and re-encoding adds a little damage. If the recording is AMR, keep the .amr as your master and produce a cleaned copy from it — the noise reducer accepts AMR directly.
Also set expectations for the source. AMR is a low-bitrate speech codec at 8 or 16 kHz. Aggressive noise reduction on an already compressed voice note can make speech sound thin or robotic. Gentle processing usually produces a more natural result than an all-out attack on the noise.
Choose a sensible strength
Most tools offer light, medium and strong settings, and the naming is genuinely useful:
- Light removes gentle hiss while keeping the voice natural. Good for recordings that are mostly clean.
- Medium is the right starting point for most noisy recordings. It balances cleanliness against naturalness.
- Strong is for material where noise is clearly drowning the voice — but it can introduce artefacts and make speech sound synthetic.
Start at medium, listen to a short sample, and only move up if the noise is still obvious. It is much easier to reduce further than to undo over-processing.
Deal with low-frequency rumble
A surprisingly common problem is rumble below the speech range: handling noise, traffic, air-conditioning units. A high-pass filter at around 80–85 Hz removes this inaudible-but-wasteful energy without affecting the voice. It is a low-risk first step that makes everything downstream work better.
The noise reducer combines a high-pass filter with frequency-domain noise reduction, so both problems are addressed in one pass. The engine runs locally in the browser, which matters if the recording is sensitive — nothing is uploaded.
Clean before you transcribe
Noise is one of the main causes of transcription errors. A model tries to make sense of every sound, and steady hum can be misheard as words — especially in quiet passages. Running a recording through the denoiser before the transcription tool can measurably reduce those mistakes.
The same goes for volume: whisper-quiet sections are where recognition tends to invent text. After denoising, use the loudness normaliser to even out the level. The two steps together often do more for accuracy than switching to a bigger model.
A note on AMR recordings specifically
AMR files deserve a mention because they are such common input for this kind of cleanup. The codec was built for clean speech over a phone network, so it already discards much of the background that a microphone might capture. That means a well-made AMR voice note often needs less noise reduction than a WAV recorded in a busy room.
What AMR does add is its own compression artefacts — a slightly watery, narrow sound. Noise reduction will not remove those, because they are part of the codec rather than the environment. If the voice is intelligible, the most useful next step is often transcription rather than more aggressive filtering. Converting the cleaned result to MP3 or FLAC then gives you a stable, shareable copy.
When to stop
The goal is intelligibility, not sterile silence. A recording with a little residual room tone usually sounds more natural and is easier to listen to for long stretches than one that has been scrubbed within an inch of its life. If the voice starts to sound underwater or robotic, you have gone too far — back off a level and accept some noise.
And if the very words are unintelligible because someone spoke over the subject or the microphone was across a busy room, no amount of filtering will recover them. In that case, a transcript with a few flagged inaudible sections is a more honest result than a heavily processed file that sounds worse.