Audacity noise reduction settings: what the three sliders actually do

Noise Reduction works by learning what your background noise sounds like and then subtracting that pattern from the whole recording. It cannot do anything until you show it a sample, and the quality of that sample decides everything that follows.

Select a stretch of audio where nobody is speaking and nothing is happening: room tone only. Two seconds is enough, four is better. Then Effect, Noise Removal and Repair, Noise Reduction, and click Get Noise Profile. The window closes and nothing appears to happen. That is correct.

Now select the whole track, open the same effect again, and this time set the sliders and press OK.

If your profile sample accidentally included a breath, a chair creak or the tail of a word, Audacity will treat that sound as noise and cut it everywhere it appears. This is the single most common reason the result sounds wrong, and no slider setting fixes it.

This is how much quieter the noise becomes. Set it to 12 dB and the background drops by roughly 12 decibels. Higher numbers remove more noise and do more damage to the voice sitting on top of it.

The default of 12 dB is a reasonable starting point and considerably lower than most people's instinct. For ordinary room noise, 6 to 12 dB is usually enough. Above about 18 dB you start hearing the processing more than the problem.

The counterintuitive part is that a little residual hiss sounds more natural than a perfectly silent background. Complete silence between words tells the ear that something has been done to the recording, even when it cannot say what.

Sensitivity decides how readily Audacity calls something noise. Low values mean it only removes sounds that closely match the profile. High values mean it removes anything vaguely similar, which increasingly includes parts of your voice.

The default of 6.00 works for most spoken recordings. If noise is still audible, raise the first slider before you touch this one. Sensitivity is where the underwater, warbling, metallic quality comes from, because consonants like s, f and t look a lot like hiss to an algorithm.

If your processed audio sounds like the speaker is in a swimming pool, this slider is almost always the cause. Bring it down before anything else.

Frequency Smoothing spreads the reduction across neighbouring frequency bands rather than treating each one in isolation. It is there to soften the artefacts the other two sliders create.

For speech, 3 is the standard value and you rarely need to move it. Zero gives you sharper noise removal with more ringing artefacts. Higher values smear the correction across a wider band, which sounds duller but less processed.

For music the calculation is different because you have real content across the whole spectrum, and heavier smoothing usually helps. For one person talking, leave it at 3 and spend your attention on Sensitivity.

For a normal room with a decent microphone: Noise Reduction 9 dB, Sensitivity 6.00, Frequency Smoothing 3. For persistent hum from a fan or air conditioning: 12 dB, 6.00, 3. For a genuinely bad recording: do not go past 18 dB, because at that point you are trading one problem for a worse one.

Use the Preview button before every OK, and preview a section with speech in it rather than a silent gap. Silence always sounds better after noise reduction. The voice is where you find out what it cost you.

The honest limit: noise reduction cannot recover a voice buried in noise. It removes a steady background from a recording that was mostly fine. If the noise is louder than the speech, or it changes over time, such as traffic or a conversation in the next room, this tool will not save it and re-recording is faster than fighting it.

If you would rather not tune three sliders per recording, Hilite does this step automatically as part of cleaning up audio. We make it, so weigh that as you like. Audacity is free and, once you know what the sliders do, entirely capable.