You recorded something important (interview, hearing, podcast, message to a client) and the audio came out with background noise. This guide shows how to remove that noise without distorting the voice — free with FFmpeg + RNNoise or with commercial software when quality matters.
Noise types and optimal technique for each
| Noise type | Example | Best technique |
|---|---|---|
| Constant | Air conditioner, computer, fan | afftdn (FFT spectral) |
| Background voices | Coffee shop conversations | RNNoise (trained on voices) |
| Intermittent | Traffic, alarms, dog barking | Manual edit + localized denoise |
| Electrical hum | 50/60 Hz power line | Specific notch filter |
| MP3 compression | Re-compressed audio multiple times | Re-encode to WAV + denoise |
| Echo | Empty-walled room | Dereverberation (ML) — FFmpeg insufficient |
FFmpeg + RNNoise recipe (best for voices)
RNNoise is trained specifically for human voices. To download and use it:
git clone https://github.com/xiph/rnnoise cd rnnoise && ./autogen.sh && ./configure && make ./examples/rnnoise_demo dirty_audio.raw clean_audio.raw
Or with FFmpeg integrated arnndn:
ffmpeg -i dirty.wav -af "arnndn=m=cb.rnnn" clean.wav
FFmpeg afftdn recipe (general use)
For non-voice-specific noise:
ffmpeg -i dirty.mp3 -af "afftdn=nr=20:nf=-25" clean.mp3
Important parameters:
nr=20— Noise reduction in dB. 12-30 typical. Higher = cleaner but more artifacts.nf=-25— Estimated noise floor. -20 (gentle) to -40 (aggressive).nt=w— Tracking mode.wfor constant,vfor variable.
Adobe Enhance Speech vs FFmpeg
Adobe Enhance Speech (part of Premiere/Audition Cloud) is brutal for cases where FFmpeg loses:
- Smartphone audios recorded outside with wind.
- Cafe interviews with conversations near same volume.
- Voices recorded with bad mic through compressed Zoom.
But costs Adobe subscription (~USD 23/month). For occasional use, contracting a service like TranscribeNode includes pre-transcription cleanup in the audio cost.
A/B validation (mandatory step)
After any denoise, compare clean file vs original on headphones (not speakers). Look for:
- Noise reduced (obvious).
- Voice does NOT sound robotic (typical over-denoising artifact).
If voice distorts, lower nr by 5 dB and re-process. Better an audio with some noise than a robotic voice.