Use a true noise-only profile and conservative reduction. Judge the voice at equal loudness, and listen to the removed signal to catch speech being mistaken for noise.
A constant fan is a better candidate than passing traffic
A laptop fan can leave a steady hiss underneath narration. Audacity’s Noise Reduction learns a profile of that background sound and uses it to distinguish likely noise from the wanted recording. It works best when the unwanted sound is reasonably consistent and the voice is clearly louder.
Passing vehicles, another person speaking and keyboard impacts change too much to fit the same profile neatly. Room reverberation is also different from steady fan noise: it is delayed speech, not a separate constant bed. Treat the specific problem you hear. More aggressive noise reduction cannot turn a distant, echo-heavy recording into a close-miked studio voice without tradeoffs, and complete silence between words is not necessarily the most natural result.
Protect the original and choose a revealing sample
Duplicate the original track or save a separate working project. Select a short passage containing quiet speech, a breath, an “s” sound and a pause. These reveal damage better than a loud, clean sentence alone. Keep a second untouched version for comparison.
Before noise reduction, inspect obvious clicks or a narrow electrical hum. Audacity’s documentation suggests a notch filter for a hum or whistle when appropriate, before broadband reduction. Do not automatically notch 50 or 60 Hz just because you know the mains frequency; confirm that the recording contains a corresponding tone. A high-pass filter can reduce low rumble, but setting its cutoff too high thins the voice. Each processor should address an audible, identified problem.
Capture a noise profile that contains no speech
- Find a pause recorded under the same fan conditions as the voice.
- Select a noise-only region. A second or two can be convenient when available, but do not include a breath, word ending, clothing rustle or keyboard tap just to make it longer.
- Open Noise Reduction and select Get Noise Profile.
- Select the short representative voice passage and reopen Noise Reduction for the actual processing stage.
A profile containing speech teaches the processor that parts of the voice are unwanted. Repeating a tiny noise sample to make a longer selection does not create new statistical information. If the fan changes speed halfway through a recording, test separate regions with profiles that match their own conditions instead of using the loudest fan section as a universal profile.
Start gently and use the Residue audition
Try a modest reduction, for example 6 dB, rather than beginning with an extreme value. Use the default sensitivity and frequency-smoothing controls as a baseline, then adjust one control at a time while previewing. These values are test settings, not a recommended preset for every microphone.
Audacity provides a Residue audition for the material being removed. Listen to it. A little ambiguous background texture can be expected, but intelligible words or prominent consonants mean the processing is taking useful speech too. Reduce the aggressiveness or reconsider the profile. Frequency smoothing can change the character of artifacts, but it does not recover voice information already removed by an earlier destructive pass. Return to the untouched copy when comparing new settings.
Recognize damage before it becomes the final voice
| What you hear | What to investigate |
|---|---|
| Metallic or watery consonants | Excessive reduction or sensitivity |
| Soft word endings disappear | Noise profile includes speech, or detector is too aggressive |
| Background swells around every word | Processing modulation or stacked gating |
| Voice becomes hollow | Broad EQ cuts, phase issues or severe overlapping noise |
Do not stack several heavy passes just because each lowers the noise further. They can also compound artifacts. A noise gate may hide pauses, but it cannot remove fan noise underneath the words. If a gate is used, a gentle expansion and sensible release often preserve more natural room sound than abrupt muting. Listen to the transition into and out of every quiet phrase.
Compare at equal loudness and in the real mix
Match the perceived voice loudness between processed and original versions before deciding which is better. A louder result often sounds clearer even when the processing itself is damaging. Switch between the same short passage at the same monitoring level.
Then listen with the intended music or other soundtrack. A small remaining noise floor may disappear in context while metallic speech stays obvious. Export a short copy in the actual delivery format and listen on headphones and a phone speaker. If the sample is still distracting, test a closer microphone position and quieter recording environment for future takes. Keep a few seconds of clean room tone in new recordings, but prioritize capturing a stronger voice signal over collecting elaborate repair presets.
Common questions
Can a noise gate remove fan noise while I am speaking?
A gate mainly reduces sound below its threshold, usually during pauses. Noise underneath speech remains and needs a different treatment.
Should the Residue preview contain only pure noise?
It is a useful diagnostic rather than a perfect separation. Clearly intelligible speech or strong consonants in the residue indicate that valuable voice material may be removed.
Sources & further reading
Written for ToolWillow with original explanations and examples. Consult current product documentation for controls that vary by version.
