Attenuate everything below a threshold by a set number of dB, with hold, hysteresis and lookahead. Draws the gain reduction over the waveform and reports what it measured.
Initializing in your browser…
Read text aloud with any voice your system provides, with sentence-by-sentence progress that works even for voices that report no word boundaries, and a reading-time estimate calibrated from the voice you pick.
Turn audio into a waveform image in five styles, exported as PNG or real vector SVG. Reads every channel, shows true level in dBFS, and redraws correctly at any size.
Find silent sections across every channel, not just the left. Threshold, hysteresis, minimum length, and cut padding all re-run instantly. Export a tightened WAV or MP3, split into clips as one ZIP, or save the timeline as CSV or JSON.
A voice track has audible room hiss in the gaps between sentences, and you want 60 dB less of it without touching the speech.
Input
vo.wav (floor -54.8 dBFS) - threshold -30 dB - range 60 dB - hold 50 ms
Output
Floor -114.8 dBFS, speech changed by -0.00 dB
Range is a level in dB, so 60 removes 60, and the passing sections come back bit for bit identical. The gain the gate applied is drawn over the waveform, and the figures above are measured from the processed audio where the gate was fully closed and fully open. The control this replaced gave about 9 dB when set to 60.
Push everything below a threshold down by a set number of decibels, and see exactly what happened: the gain reduction the gate applied is drawn over the waveform, and the panel reports how far the noise floor moved and by how little the wanted signal was touched. Range means what it says, so 40 dB removes 40 dB.
Reduce room tone during pauses so only the speaker is heard, without touching the speech.
Gate individual drum microphones to reduce bleed from other kit pieces.
Mute room noise in recorded segments where nobody is speaking.
Remove air conditioning hum and ambient sound between lines of narration.
The gate follows the level of the loudest channel through a sliding detector window, opens when that level crosses the threshold, holds open while it stays there and for a set hold time afterwards, then closes over the release. Attack, hold and release are travel times across the whole range, so changing how much attenuation you want does not quietly change how long the gate takes to open. Hysteresis sets how far below the threshold the level has to fall before the gate is allowed to close, which is what stops material sitting right at the threshold from flickering. Lookahead brings the decision forward by a few milliseconds so the gate is already open when the start of a word arrives rather than clipping it.
Three things were measurably wrong with the previous build, and each was measured before it was touched. Range was a linear gain fraction rather than a level, so asking for 60 attenuated the noise floor by 8.99 dB and almost all the useful travel was crammed into the top few percent of the slider; it now hits 12, 24, 40, 60 and 90 dB to within 0.05 dB of the number shown. The attack ramp ran at a fixed rate across a distance that depended on Range, so an attack set to 20 ms measured 10.5 ms; both slopes now take the time they are set to whatever the range is. And each channel was gated from its own level, so a burst on the left pushed the right channel down 9 dB and dragged the stereo image sideways; the detector is now linked across channels by default, with an independent mode for when that is what you want.
Material above the threshold is left alone. Measured on a burst fixture, the passing sections come out bit for bit identical to the input, a change of 0.00 dB, and a steady tone ten decibels above the threshold shows no added modulation and no added distortion. Cross-checked against ffmpeg's agate on the same file, the gate agrees about when it should be open across 98 percent of the audio, with a mean difference of 0.50 dB.
The red curve over the waveform is the gain the gate actually applied, sample by sample, so you can see where it closed and check that against the gaps you meant it to close in. Under it, the Measured Result panel compares the audio before and after: the floor is read where the gate was fully closed and the signal where it was fully open, with the ramps between them excluded from both, so neither number is muddied by the transitions. Gating runs in a Web Worker, so the tab stays responsive on long files, and export writes a 16-bit WAV or a 192 kbps MP3.
Set it just above the noise floor and below the quietest thing you want to keep. The shaded band on the waveform shows where the threshold sits in the same amplitude space as the audio, so you can see which parts fall inside it.
It is how far the quiet parts are pushed down, in decibels. Set it to 40 and the noise floor drops by 40 dB. Full mute is available at the top of the range. The previous build treated the control as a linear gain fraction, so 60 gave about 9 dB of reduction.
No. A gate only acts when the level is below the threshold, so noise underneath speech stays. It removes noise in the gaps, which is usually where it is most noticeable. Broadband noise reduction is a different problem.
A fast attack around 1 to 5 ms catches word starts, and a release of 100 to 250 ms with 50 to 150 ms of hold lets tails decay naturally instead of being chopped. The Voice preset is set up that way.
It should not, because the detector is linked across channels by default: one decision drives both sides. If you deliberately switch to independent mode, each channel gates itself and the image can shift when the two sides differ.
Audio is decoded and processed locally with the Web Audio API. Your files are never uploaded to a server.