Ambient noise is the reason a technically fine recording still sounds amateur: the laptop fan, the HVAC hum, traffic through a window, the hiss on a phone voice memo. Audacity can clean up most of it, but only if you match the tool to the type of noise you have. Here is the exact process, the settings that matter, and where manual editing stops being worth your time.

Step 0: Identify the noise before you touch a slider

Audacity's main noise tool assumes the noise is stationary — a steady fan, hiss, or hum that sounds about the same in the quiet gaps as it does under your voice. If the noise changes from moment to moment, noise reduction will smear it around instead of removing it.

Method 1: Noise Reduction, the workhorse

This is the effect you will reach for 80% of the time. Do it before any EQ or compression, because compression raises the noise floor.

  1. Find a stretch of noise only, at least half a second and ideally 1–2 seconds, taken from a gap between sentences. Do not include a breath, a lip smack, or the tail of a word.
  2. Select that region with the mouse.
  3. In Audacity 3.4 and later, go to Effect → Noise Removal and Repair → Noise Reduction. In older versions it is Effect → Noise Reduction. Click Get Noise Profile. The window closes; that is expected.
  4. Select the entire track (Ctrl+A) and reopen the same effect.
  5. Start with Noise reduction 8 dB, Sensitivity 6.00, Frequency smoothing 3.
  6. Hit Preview on your loudest sentence. Listen for a metallic, watery, or swishing tail on the ends of words. If you hear it, drop the reduction to 6 dB.
  7. Apply. If noise remains, run a second pass at 4–6 dB rather than one pass at 18 dB.

Practical ceiling: most speech tolerates 10–12 dB of reduction before it sounds processed. Past about 15 dB, voices develop that underwater quality that is impossible to fully undo.

Why two gentle passes beat one aggressive pass

Each pass removes noise but also chips away at consonants and breath detail. Two passes at 6 dB leave the voice sounding more natural than one pass at 12 dB, even though the measured noise floor ends up similar. It costs you an extra minute and saves the take.

Method 2: High-pass and notch filters

Noise reduction does not handle low-frequency rumble or electrical hum well. These two filters do, and they are fast.

  1. High-pass filter: select the track, then Effect → EQ and Filters → High-Pass Filter. Set 80 Hz for a deep male voice, 100 Hz for most other voices, rolloff 12 dB per octave (24 dB if the rumble is severe). This kills HVAC drone, footsteps, and desk bumps without touching intelligibility.
  2. Notch filter for hum: if you hear a steady buzz, use Effect → EQ and Filters → Notch Filter. Set frequency to 60 Hz in North America or 50 Hz in Europe and most of Asia, Q to 10. Then repeat at the harmonics — 120, 180, and 240 Hz (or 100, 200, 300) — because hum is almost never a single tone.

Method 3: Gate the gaps, not the speech

A noise gate turns down the signal when you are not talking. It does not remove noise underneath your voice, so it is a finishing touch, not a fix. If you want quieter gaps, select a silent region, go to Effect → Volume and Compression → Amplify, and set it to around -15 dB with clipping disallowed. Repeat on each gap, or use Effect → Special → Truncate Silence for long dead air. Keep gate thresholds conservative; aggressive gating produces audible pumping that is more distracting than the original hiss.

Method 4: Repair one-off noises

For a single mouse click, cough, or dog bark, do not run noise reduction over the whole file. Zoom in, select just the offending 100–300 milliseconds, and use Effect → Noise Removal and Repair → Spectral Delete, or the Spectral edit tools from the track's dropdown menu. The Repair effect handles clicks shorter than about 128 samples. For vinyl-style crackle, Click Removal helps.

Where manual editing runs out of road

When AI voice isolation is the better call

AI voice isolation models are trained to separate speech from everything else, so they handle the cases Audacity cannot: variable noise, reverb, music beds, keyboard noise, crowded rooms. Instead of profiling a noise sample, you upload the file and the model reconstructs the voice. There is nothing to install — a free tool like ClearVoiceHQ (clearvoicehq.com) takes a file straight in the browser, which is handy when the recording is a Zoom call or a phone memo you want fixed in two minutes rather than twenty.

The tradeoff is control. AI isolation can slightly dull sibilance, soften breath sounds, or leave faint gurgling on very quiet passages, and you cannot dial in a specific frequency the way you can with a notch filter. It is also less predictable across files.

Pick your path

  1. Steady fan or hiss, and you have time: high-pass filter, then Noise Reduction at 8 dB. Best quality per minute spent is usually here.
  2. Hum only: notch at 50/60 Hz plus harmonics. Done in under a minute.
  3. Variable noise, music, or a second voice: run it through ClearVoice first, then use Audacity only for trimming, high-pass filtering, and loudness normalization to around -16 LUFS for podcast or -14 LUFS for video.
  4. Hybrid, for important recordings: AI isolation first, then a single light Noise Reduction pass at 4 dB in Audacity to catch residual hiss, then a high-pass filter at 90 Hz.

The honest summary: Audacity is still the better tool when the noise is simple, steady, and you want maximum control over the result. AI voice isolation wins when the noise is messy, the deadline is short, or the recording came from a laptop mic in a room you do not control.