What audio noise reduction can and cannot fix

Identify the problem before applying a filter

Listen to a quiet gap and to normal speech. A steady fan or background hiss is a different problem from a microphone rubbing against clothing. A sudden bump needs local attention; room reverberation is the voice repeating through the room; clipping is distortion already captured in the signal.

The free noise reducer applies signal processing. It does not use a reference recording of clean speech and does not promise to reconstruct lost syllables. Its strength setting is a trade-off: more suppression can also remove useful voice detail.

Before processing, fix anything you can at the source. Turn off an avoidable appliance, move closer to the microphone, reduce reflections or lower gain when loud words distort. A better short test recording is often more reliable than aggressive cleanup of the full session.

Make a restrained comparison

Do not use silence in the result as the only success criterion. A gate can make pauses quiet while leaving noise under every word. The listener hears the speech, not just the gaps between it.

1. Choose an untreated source and keep it unchanged. Listen to a passage with quiet and loud speech, pauses and difficult consonants.
2. Open the noise reducer and choose a light treatment first. Run the export and download the result.
3. Compare the same passage at similar listening volume. A louder version can seem better even when its speech detail is worse.
4. Listen particularly to word endings, breath sounds and the transition into pauses. Reduce the treatment if the voice becomes watery, metallic or dull.
5. If the result helps, listen through the whole recording. A setting that works on the first speaker may damage a quieter guest.

Know when to stop processing

If no light setting improves the voice without audible damage, keep the source or use more specialised editing. Do not label a worse result “clean” because the numerical noise level is lower. A measurement helps locate a change; listening establishes whether that change serves the recording.

• Compare consonants and short words with the original. A quieter background is not worth losing the meaning of the sentence.
• Listen for pumping where the background rises and falls with speech.
• Check whether different speakers remain equally understandable. A quieter voice may be suppressed more than the main speaker.
• Check music and intentional ambience. A speech-oriented treatment may change their character.
• Listen on headphones and a small speaker. Each can reveal different problems, especially in low-level passages.

Normalize only after you know what remains

Raising the level after cleanup may make remaining noise more obvious. That does not necessarily mean normalization introduced noise; it may have raised what was already there.

Keep cleanup and level adjustment as separate review decisions. Noise reduction addresses unwanted sound; normalization addresses overall loudness and peak constraints. Neither replaces microphone placement or a fair edit.

When the recording contains private information or an unusable interruption, trim that passage deliberately. Do not expect a cleanup filter to know what should be removed. The final review should cover intelligibility, meaning and continuity, not just a processing status.

Questions about this workflow