Audio Editor

Two things, both of which are normally found only in paid desktop software. The first is spectral repair: the spectrogram is an editing surface, so you can draw a box around a single unwanted sound and remove it without touching the audio around it. The second is proper loudness metering and mastering — integrated LUFS, loudness range, and true peak measured to the ITU-R BS.1770 standard, with one-click mastering to the published targets for Spotify, YouTube, Apple, and EBU R128 broadcast. Most free online editors offer trimming and a volume slider; neither of these is in that category. Everything still runs in the browser, so nothing is uploaded.

A spectrogram shows the recording as frequency against time, so a sound occupies a patch rather than a moment. A phone notification, a chair squeak, a bird outside the window or a single cough sits in a small region of that picture, while the voice underneath occupies different frequencies at the same instant. Selecting that patch and removing it takes out the interruption and leaves the speech. Cutting on the waveform would remove the speech as well, and an EQ notch would hollow out those frequencies for the entire recording to fix one second of it.

Peak level tells you the highest sample in a file. It says almost nothing about how loud that file sounds — a heavily compressed track and a dynamic one can both peak at exactly 0 dBFS while one is obviously much louder. LUFS measures perceived loudness instead, by filtering the signal the way the ear weights frequency and averaging over time with the quiet parts gated out. Every streaming service and every broadcaster now specifies a LUFS target, so it is the number that decides whether your file plays at the right volume.

Because the service turns it back down. Spotify, YouTube and Apple all normalise to their target on playback, so a master delivered at −8 LUFS against a −14 target is simply attenuated by six decibels. What does not come back is the dynamic range that was crushed to reach −8 in the first place. The result is a track that plays at the same volume as everyone else but sounds flatter than it needed to. Hitting the target is not a compromise; it is the whole point.

Sample peak is the largest value stored in the file. True peak is the largest value the waveform actually reaches between those stored samples, once a converter or a lossy encoder reconstructs the curve through them. A file reading exactly 0.0 dBFS at every sample can still overshoot above full scale in between and distort on playback — which is why delivery specifications ask for a ceiling of −1 dBTP rather than 0. The meter here estimates true peak by reconstructing between the samples, and the limiter works to that figure rather than to sample peak.

Because limiting changes the thing it was applied to protect. Applying the gain needed to reach a target and then limiting the peaks removes energy, which lowers the loudness — so a single pass lands consistently short of where it aimed. The tool measures again after limiting, corrects the residual error, and reports what it actually achieved rather than what it intended. If a tool tells you the target without measuring the result, it is telling you its intention.

Use the Deeper preset, or move both sliders down together. The mistake that produces a monster is moving pitch on its own: pitch is only half of what tells you whose voice you are hearing, and the other half is the formants — the fixed resonances of the throat and mouth, which do not move when you sing a lower note. Drop pitch alone and you get a slowed-down tape. Drop the formants with it, by a smaller amount, and you get a larger person. Every preset here moves the two by different amounts for that reason.

The vocal folds make a buzz at a pitch; the throat and mouth then act as a resonant tube that reinforces certain frequencies and suppresses others. Those reinforced bands are the formants, and they depend on the size and shape of the tube rather than on the note. A child does not sound like a child because their pitch is high — it is because their vocal tract is short, which puts the formants high. That is why a pitch shifter alone cannot make an adult sound like a child, and why this tool gives the two separate controls.

No. Pitch is moved by a phase vocoder, which stretches the recording in time without touching the pitch, after which resampling puts the length back and leaves the pitch moved. Speed and pitch stay independent, so a lower voice speaks at exactly the same pace.

Each runs a whole workflow in the order an engineer would: clean the noise, tame the sibilance, shape the tone, control the dynamics, then set the delivery loudness last. Last matters — every stage before it changes the loudness, so measuring earlier would be measuring something that no longer exists. Podcast finishes at −16 LUFS, Broadcast at −23, Audiobook at −20 and Commercial voiceover at −16, each with the matching true-peak ceiling.

No. The file is decoded and every operation runs in your browser. Nothing is transmitted to any server at any point, there is no account, and closing the tab is the end of it. For interviews, medical recordings, unreleased music and confidential meetings, that is the difference between usable and not.

An EQ turned down at 6 kHz removes the harshness of an "s" and takes all the air and detail with it, for the whole recording. The de-esser looks only at the sibilant band, frame by frame, and turns it down only in the frames where it is genuinely too loud. Everywhere the speaker was not hissing, the brightness is untouched.

It finds every silence longer than a second and shortens it to about a third of a second, rather than removing it. Deleting pauses outright is what makes edited speech sound breathless and wrong — the rhythm of talking is partly the gaps. Trimming them keeps the beat and loses the dead air, which is usually what you actually wanted.

Yes, up to twenty steps, on every operation including the processing ones. The history keeps whole copies of the audio rather than trying to reverse each operation, which uses more memory but cannot drift out of step with what you are looking at — and drift is the failure mode that loses work.

None imposed, since nothing is uploaded. The practical limit is memory: decoding produces raw samples at roughly ten megabytes per minute of stereo CD-quality audio, and the undo history keeps copies. An hour of audio is comfortable on a normal computer; phones have far less room.

For a single file — editing, repairing, cleaning and mastering it to a delivery target — it does the jobs that matter and does them properly. What it is not is a multitrack workstation: there is no mixing several recordings together, no MIDI, no plugins and no automation. If you need to assemble a session from many sources, use a DAW. If you need to take one recording and make it deliverable, this does that without installing anything or uploading your audio.