Key takeaways
- A noise gate mutes the pauses but leaves the noise under the words, and a tight one chops word endings and breaths.
- A low-cut (high-pass) filter removes the rumble below the voice, around 80 Hz, but EQ can’t separate noise and voice that share frequencies.
- Profile-based spectral reduction needs a noise-only sample and suits steady noise; neural noise suppression adapts on its own, every 10 ms or so, to noise that changes.
- Echo and other voices are separate problems: de-reverb for the first, nothing reliable for the second.
The short answer
| Tool | How it works | Good for | Fails at |
|---|---|---|---|
| Noise gate | Mutes the signal whenever it falls below a threshold | Silencing faint noise in the pauses of a clean recording | Noise under the words stays; quiet word endings and breaths get chopped |
| EQ, low-cut, high-pass | Turns chosen frequency ranges down | Rumble, wind thumps, the lowest note of mains hum | Noise that shares the voice’s frequencies: fans, hiss, traffic, voices |
| Spectral noise reduction (noise profile) | Learns the noise from a noise-only sample, then subtracts it everywhere | Steady hiss, hum and fans in long recordings | Noise that changes; needs a clean sample; watery artifacts when pushed |
| Neural noise suppression | A trained model tells speech from noise, every 10 ms or so | Steady and changing noise under and between the words, with no sample | Echo, other voices, songs with vocals; can thin a voice when pushed |
| De-reverb | Estimates the room’s reflections and turns them down | Echoey, distant voices | Background noise; sounds processed when pushed hard |
Noise gate
A noise gate is an automatic mute. When the level falls below a threshold you set, it closes and silences the signal, or turns it down by a set range; when the voice rises above the threshold, it opens. Attack, hold and release settings decide how fast it reacts.
It tidies the pauses of a recording that is already decent: a little hum between sentences, a quiet room. It does nothing while you speak, because the noise is still there under every word, so the background pumps in and out with the voice, which is often more distracting than steady noise. Set too high or too fast, a gate also cuts the quiet ends of words, breaths and soft consonants. An expander is its gentler cousin: it turns quiet passages down progressively instead of shutting them.
EQ, low-cut and high-pass
An equalizer turns frequency ranges up or down. A low-cut, also called a high-pass filter, removes what lies below a chosen frequency, and it is the one EQ move almost every voice recording benefits from: the fundamental of most speaking voices sits above about 85 Hz, while traffic rumble, air conditioning, handling noise and wind thumps sit largely below it. A cut around 80 Hz takes them out and leaves the voice essentially intact.
Mains hum is a special case. Its fundamental, 50 or 60 Hz depending on the country, sits below the cut and is turned down with the rumble, but its harmonics, at multiples of that frequency, run up into the voice, where only narrow notch filters or noise reduction can reach them (see electrical hum). That is the limit of EQ in general: when the noise and the voice share frequencies, as fans, hiss, traffic and other voices do, cutting the noise cuts the voice too.
Spectral noise reduction from a noise profile
The classic noise-reduction tool in desktop audio editors works in two steps. You select a passage where nobody speaks, and the tool learns the noise’s average spectrum: its profile, or print. It then splits the whole recording into short slices of frequencies and turns down, in each one, whatever does not rise above that profile.
It is effective on steady noise in long recordings: tape hiss, a constant fan, a hum. It needs a clean, noise-only sample, and it assumes the noise never changes, so traffic, wind, keyboard clicks or a fan that speeds up slip through or come out half-removed. Pushed hard, it leaves musical noise, the watery, warbling texture described in why noise reduction sounds robotic.
Neural noise suppression
Neural noise suppression uses a model trained on large amounts of speech and noise to recognise, moment by moment, which parts of the sound are voice. RNNoise, the open-source model from Xiph.Org that Noise Remover runs, looks at the sound in 10 ms slices at 48 kHz, splits each slice into 32 frequency bands and decides a gain for every band, along with an estimate of whether someone is speaking. It needs no noise sample, and it adapts as the noise changes.
That is what makes it work on changing noise. In our lab test, keyboard clicks dropped by 11 to 13 dB, traffic by 24 to 37 dB and wind by 19 to 27 dB, with the voice within about 1 dB of its level on most noises. Its limits are the flip side of its training: it keeps whatever sounds like speech, so other voices and songs with vocals stay, overlapping voices can pull the main voice down a little, and echo, the voice itself reflected, is left in place. How we measured.
| Noise | Noise removed during pauses, in dB | Change in voice level |
|---|---|---|
| Electrical hum | 24 to 41 dB | −0.3 to −1.6 dB |
| Traffic | 24 to 37 dB | −0.1 to −0.3 dB |
| Wind gusts | 19 to 27 dB | −0.3 to −1.3 dB |
| Keyboard typing | 11 to 13 dB | 0.0 to −0.2 dB |
| Not what it is made for | ||
| Another voice | 0 to 2 dB | +0.2 to −0.6 dB |
De-reverb: a different problem
Room echo is not noise. Reverberation is your own voice arriving again from the walls, a few milliseconds to a second or so late, so to any noise tool it looks like speech. De-reverb tools, found in some desktop audio editors and plugins, estimate the room’s tail and turn it down; they handle moderate echo well and sound processed when pushed. Gates and noise reduction can’t do this job. More in echo or background noise?.
In what order to use them
Tools in a chain affect each other, so their order matters. For a voice, a common order is:
- Low-cut, to clear the rumble before anything reacts to it.
- Noise reduction, and de-reverb if the room needs it.
- EQ, to shape the tone of the voice once the noise is gone.
- Compression, to even out loud and quiet passages.
- Loudness normalization, or a limiter, last.
What Noise Remover combines
Noise Remover chains the tools that suit a voice recorded on a phone, in that order:
- A low-cut at 80 Hz, a high-pass filter applied before anything else, for the rumble below the voice.
- RNNoise, the neural suppression, which removes steady and changing noise under and between the words.
- An intensity cap: the original, delayed by the engine’s 20 ms so the two line up, is mixed back in far below the cleaned voice, which limits the reduction to a cap you choose, about 29 dB by default.
- Loudness normalization: one gain, measured on the final mix, to reach −16 LUFS unless the −1 dBTP ceiling stops it first.
There is no gate, because the cap and the neural model handle the pauses without chopping words, and no EQ beyond the low-cut, which leaves the voice’s own tone alone. The whole chain runs on the phone, and nothing is uploaded. How it works shows each stage.
Hear the whole chain on your own recording.
Free · Android 8.0 or later, 64-bit
When a desktop audio editor is still the right tool
- Music. Speech models are trained to keep voices and treat instruments as noise; music needs tools made for it.
- Multitrack recordings. A podcast with one microphone per person is cleaned track by track, often with a gate or an expander on each.
- Echo. De-reverb is a desktop job.
- Heavy restoration. Clicks and crackle on old tapes, clipped peaks, a phone ringing in the middle of a sentence: spectral editing and repair modules do what no one-pass cleaner can.
- A strict delivery specification. Hitting an exact loudness with a limiter, or shaping the voice with EQ and compression, is mixing work.
For a single voice recorded on a phone (a voice-over, a podcast take, a lecture, a voice note), a one-pass cleaner that runs on the phone covers most of what is needed.