Some noise comes out of a recording almost perfectly. Some cannot be removed at all, and trying leaves the voice sounding like it is underwater. The difference is not about how good the software is — it is about what kind of noise you have.
The distinction that decides everything
Steady, predictable noise is easy. Air conditioning, computer fans, tape hiss, mains hum, distant traffic. These sit at consistent frequencies and never really change, so the software can learn what they look like and subtract them. Results here are often dramatic.
Unpredictable noise is hard. A door slamming, a dog barking, someone coughing, another conversation. These occupy the same frequencies as speech and arrive without warning. There is nothing consistent to model, and removing them means removing parts of the voice too.
Choosing a method
Different noise wants different treatment, which is why there is a choice rather than one button:
- AI Noise Removal – the best option for speech. It has learned what voices sound like and keeps them while stripping the rest. Start here for interviews, podcasts and voice memos.
- Spectral Gate – also good for speech. Builds a profile of the noise and subtracts it. Effective on steady background noise.
- Wiener Filter – fast and general-purpose. A sensible first try when you are not sure what you are dealing with.
- Median Filter – targets clicks and pops rather than continuous noise. The right choice for vinyl crackle or digital dropouts.
- HPSS + Median – separates harmonic from percussive content, which suits music with vocals rather than plain speech.
Use the preview before committing. Denoising is a judgement call about an acceptable trade-off, and hearing it is the only reliable way to make that judgement.
Why over-processing sounds worse than noise
Every denoising method works by deciding which parts of the signal are noise and removing them. Push it too hard and it starts removing parts of the voice as well — particularly the quiet, high-frequency components that make consonants intelligible.
The result is the distinctive artefact people describe as watery, metallic, or underwater. Words become harder to make out even though the recording is objectively "cleaner". A listener will tolerate a noticeable hiss behind a clear voice far more readily than a muffled voice in silence.
Getting it right at the source
No processing beats not recording the noise in the first place, and the fixes are mundane:
- Get the microphone closer. Halving the distance dramatically improves the ratio of voice to room noise. This matters more than equipment.
- Turn things off. Fans, air conditioning, fridges, a laptop working hard.
- Choose a soft room. Carpet, curtains and furniture absorb reflections. Bare walls and glass create echo, which no denoiser can remove — reverb is your own voice arriving late, and it cannot be separated from the direct sound.
- Record a few seconds of silence. A clean sample of the room tone gives noise-profiling methods something accurate to work from.
A sensible order of operations
- Trim first. Cut out the parts you do not need, including one-off bangs and interruptions. No point processing audio you are going to discard.
- Denoise second, gently, checking the preview.
- Adjust volume last. Raising the level first would amplify the noise before you remove it.
Common questions
Can I remove someone talking in the background?
Not really. Another voice occupies exactly the same frequency range as the one you want, and no filter can reliably tell them apart. If the interruptions are brief, cutting those sections out is the practical answer.
Can I remove echo from a room?
No. Echo is your own voice reflected back a fraction of a second later, so it is not separable from the signal. It has to be prevented at recording time with soft furnishings and a closer microphone.
Why does my voice sound robotic afterwards?
Too much reduction. Ease off, or switch to AI Noise Removal, which is better at preserving speech than the more mechanical filters.
Should I denoise before transcribing?
Yes, gently. Cleaner audio measurably improves speech-to-text accuracy. Just avoid heavy processing, since the artefacts it introduces confuse recognition as much as the original noise did.