Audio Editor

How it works

Every file you add becomes a clip on a track. A clip is a window onto the original audio, not a copy of it: when you trim, split or move one, only its start and length change, and the samples underneath stay exactly as they arrived. That is what "non-destructive" means, and it is why a trim can be pulled back out again later and why undo goes back two hundred steps without eating memory.

Effects come in two kinds. Track effects run live on a track as it plays and never touch the audio, so you can change them at any point. Process effects, like noise removal or a loudness target, write a processed copy and point the clip at it, so undo still takes them off. The project saves itself in your browser as you work; close the tab, come back tomorrow, and the editor offers to pick up where you left off. To move a project to another computer, the project menu saves it as a single .audiosaw file, which is an ordinary zip of your audio and a JSON description of the timeline.

Jobs it is good at

• Cutting a section out of the middle. Drag along the ruler over the part you do not want and press Delete. The gap closes up, the way deleting a sentence does. The single-purpose audio cutter cannot do this; it keeps one region.
• A podcast or voice-over with music. Voice on one track, the music bed on another. Load the Podcast voice patch on the voice and Music under voice on the bed, pick the voice track in its ducker, and the music dips by itself whenever someone talks. Put Podcast master on the master and export as MP3.
• Recording over a backing track. Press Record and the other tracks play while you sing or talk; the take lands on its own track, lined up. Starting from nothing, turn on the click and a one-bar count-in and play to the metronome. Record several passes over a loop, then keep the best phrase of each.
• Mashups, edits and ringtones. Line up two songs, split at the beat, duplicate a chorus, loop a section to check the join, then export just the selected range.

Recording and keeping it in time

Most browser recorders use MediaRecorder, which hands back a compressed file with no timing information, so a take recorded over a backing track has to be lined up by eye. Here the microphone is read sample by sample on the same audio clock the playback runs on, and the delay through your speakers and microphone, as the browser reports it, is taken off when the take is placed. Two things still matter. Wear headphones, or the backing track bleeds into the microphone and ends up in the take. And Bluetooth headphones add a delay that browsers do not always report accurately; if a take sounds late, select it and press ← a few times. Each press moves it 10 ms.

Echo cancellation, noise suppression and automatic gain are switched off for recording. They are built for video calls and damage music: automatic gain pumps the level up and down, and noise suppression treats a held note as noise.

The mixer: live effects on every track

Press FX on a track, the Mixer button, or the F key. Every track has its own effect chain, and so do the master and any number of return buses. Effects run while the song plays and you hear a change the moment you make it: nothing is rendered, the original audio is never altered, and every knob turn is one step of undo. The mixer docks at the side on a laptop and slides up from the bottom on a phone, with the timeline still usable above it, so you can scrub, loop and tweak at the same time.

There are three ways in, depending on how much control you want:

• Patches. A whole chain in one tap. There are twenty, including Podcast voice, Radio DJ, Singer, polished, Telephone, Walkie-talkie, Robot, Lo-fi, Underwater, Punchy drums, Fat bass, Guitar amp, Dub delay and, for the master, Streaming master and Loud master. You can save your own chain as a patch.
• Smart controls. Each effect opens on two to four plain-words knobs, and each knob moves several settings together. The compressor's Amount lowers the threshold and raises the ratio as you turn it. The reverb's Space lengthens the decay, adds pre-delay and pulls back the early reflections, the way a bigger room would. The EQ has Bass, Clear the mud, Presence and Air.
• Show all controls. Every parameter, grouped, with a slider and a value you can type. The EQ has a low cut and a high cut at 12, 24 or 48 dB per octave, two shelves and four bands, which you can drag directly on its curve with the live spectrum behind it. The compressor draws its transfer curve and shows gain reduction as it happens.

Every effect has an on/off switch, an output level and, where it makes sense, a Mix control for blending the dry and processed sound. A/B lets you flip between two settings of the same effect to compare them. There are 89 factory presets across the 25 effects, and you can save your own.

The 25 are grouped the way a studio would group them:

• EQ and filter: Channel EQ, a resonant filter with an LFO that can sync to the tempo (from a slow build to a dubstep wobble), and an exciter.
• Dynamics: compressor, look-ahead limiter, noise gate, de-esser, three-band multiband compressor, transient shaper, and a sidechain ducker that turns one track down whenever another plays.
• Colour: drive with six characters from tube to wavefolder, bitcrusher, and tape saturation with wow and flutter.
• Modulation: chorus, flanger, phaser with 2 to 12 stages, tremolo, auto-pan and ring modulator. The Robot voice preset is the classic 30 Hz ring-mod Dalek.
• Delay and reverb: a stereo delay with ping-pong and filters in the feedback path (the Dotted 8th preset is the Edge's guitar trick), a tape echo in the spirit of a Roland Space Echo that saturates as it runs away, and a reverb generated from its settings.
• Pitch and utility: a real-time pitch shifter for harmonies and doubling, a utility for gain, stereo width, balance, polarity, channel swap and mono, and an analyzer.

Sends, buses and the tempo

Every project starts with two return buses, Reverb and Delay. A track's Send knob copies it into a bus, so a vocal and a guitar can share one room and sound like they were recorded together. The effects on a bus are 100% wet because the dry sound is already on the track. You can add more buses and put anything on them. Synced delays, tremolos, auto-pans and filter sweeps follow the project tempo. Type it in, tap along, or let the editor read it from the audio with the same detector the BPM finder uses.

Bars, beats and a metronome

Click the tempo in the corner of the ruler, or press G, and the ruler counts bars and beats instead of minutes and seconds. Bar lines are drawn heavier than the beats between them, and as you zoom in the grid divides down to eighths and sixteenths. The clock beside the play button reads bar, beat and sixteenth, with the time underneath. With the bars ruler on, anything you drag lands on the grid: clip moves, trims, range selections and markers. It still jumps to a nearby clip edge, marker or the playhead first, so two clips butt together cleanly. Hold Alt to drag freely. Double-click the ruler to select the bar under the pointer, then press L to loop it, which is how you practise a riff or check an edit's join four bars at a time.

The tempo sheet sets the time signature: 2/4 up to 7/4, and 3/8 up to 12/8. The tempo always counts quarter notes, as it does in Logic, Ableton Live and GarageBand, so 6/8 at 120 clicks six eighth notes a bar, 0.25 s apart. Plenty of recordings do not start on a downbeat at exactly 0:00. Put the playhead on the first strong beat and press Bar 1 at the playhead, and the whole grid shifts to match, including the bars before it.

Press K, or the Click button in the toolbar, for a metronome. The downbeat is a higher tick than the other beats, and in 6/8, 9/8 and 12/8 the start of each group of three gets a middle one. The clicks are scheduled on the same audio clock as the clips, a fraction of a second ahead, not fired by a timer. On every release a headless browser plays the timeline and checks that consecutive clicks are exactly one beat apart on that clock. The click goes straight to your speakers, past the master effects and the meters. The same check exports a silent project with the click switched on and confirms the export's peak is exactly zero, so the click cannot leak into an export, a stem or anything sent to a tool.

A count-in of one or two bars plays before recording starts, and a large number counts the beats down. Recording actually begins on the downbeat: the take is placed at the playhead, and the count-in audio the microphone heard is dropped instead of being written over whatever sits before it on the track. Wear headphones with the click on, or the microphone records it.

Loop recording and comping takes

Select a range, turn on Loop (L) and press Record. The editor plays the range over and over, and you play or sing along each time round. When you stop, every pass is kept. The last full one goes on the track, and the earlier ones are stored as takes of that stretch, along with whatever was on the track there before you started. A pass cut short in its first half is dropped as an accident. It is all one recording underneath, cut at the moments each pass started on the audio clock, so every take lines up with the backing exactly as the single-take placement does.

A clip with takes behind it shows a count in its corner. Right-click it, or hold it on a phone, and choose Takes…. Listen to each and press Use. To comp, keeping the best phrase of each take, select a range first: Use then replaces only that range. Choosing a take swaps it rather than deleting anything. What was playing moves into that take, so you can always swap back, and each seam gets a 5 ms fade so it does not click. The check that runs on every release confirms a swap loses no audio and a second swap restores the original.

The loop goes round without a gap. Each pass is scheduled on the audio clock a moment before the one before it ends, so the join is sample-exact, and the check that runs on every release confirms the passes start exactly one bar apart. Takes travel with their clip: move it, even to another track, and its takes go too. Delete it and they go with it, and a ripple delete, an inserted gap or a crop shifts them along with the audio.

Fitting a loop or a sample to the song

Most loops and samples arrive at the wrong tempo and in the wrong key. Open the tempo sheet (click the tempo in the ruler's corner) and press Detect from the audio. The tempo is read from the longest clip, or the one you selected, and the grid lines up with its beats as well: bar lines land on the kicks, not wherever 0:00 happened to be. The beat position is measured. On synthesized drum patterns from 84 to 174 BPM, the first beat landed within 1.3 ms, and within 3.1 ms on strummed chords with no drums; the check that runs on every release allows 10 ms. Which beat is one is a guess made from where the chords change, where the kick and bass land, where the snare's backbeat falls and where a crash rings. It was right on all 22 test patterns (rock, dance, a waltz, strumming, and drums with a crash on one, each starting partway into a bar), but real songs can fool it, so the sheet tells you where it put bar 1 and how to move it. Detect the key in the same sheet reads the project's key with the engine behind the key finder.

Then right-click a clip, or hold it on a phone, and choose Fit to the project tempo. The clip's own tempo is detected and turned into the half or double reading nearest the project, so a 70 BPM loop going into a 140 project is treated as a 140 loop. It is then time-stretched with ffmpeg without changing its pitch. The result is held to exactly the right length: a 19.5-second loop at 100 BPM comes back exactly 16.25 seconds long at 120, not a few milliseconds over. A loop that is even slightly too long drifts off the grid a little more on every repeat.

Match the project key reads the clip's key and offers the shortest shift into the project's key, up or down. A minor-key clip counts as already fitting its relative major, since the two share every note. It also offers the other direction (up 5 or down 7 land on the same notes an octave apart) and a button for when the key was misread. In testing, an E minor loop moved up five semitones into an A minor project and then read as A minor. Both fits are one step of undo, like every other Process effect.

Slice at the hits splits a drum loop into one clip per hit, with the detector behind the sample slicer. Each cut sits just before its attack, on a zero crossing, so the pieces can be moved, duplicated and rearranged on the grid without clicks. It is one step of undo.

Detect chords, in the same menu, writes the chords along the bottom of the clip, and they stay with the audio when you trim or move it. With the bars ruler on, chord changes land on the project's beats. The detector is the one behind the chord finder, and that page lists what it names and how it was measured.

MIDI tracks

Drop a .mid file on the editor and each part in it becomes a track of notes. It reads files from other programs as well as its own: several tracks or one, tempo changes partway through, the sustain pedal (written into the note lengths), and General MIDI drums on channel 10, which play on a drum kit. An empty project takes the file's tempo and time signature, so the bars line up with the music. A MIDI clip shows its notes instead of a waveform. It is an ordinary clip: it splits, trims, moves, loops, fades and undoes like any other, runs through the track's effects and automation, and is part of every export.

Right-click a MIDI clip, or hold it on a phone, for its menu. Edit notes… opens the same piano roll as audio to MIDI, on the project's grid, with Play to hear the changes before you save them. Instrument… chooses what the track plays: electric keys, a warm pad, a pluck, synth bass, a lead, an organ or the drum kit. They are simple synth voices built from oscillators and filters, not recordings of real instruments, and they are what you hear in the export. A part whose file names an instrument (General MIDI program numbers) starts on the nearest voice. Download .mid writes the clip's notes back out for a DAW, at the project tempo.

Convert to MIDI, on an audio clip's menu, listens to a single voice or instrument, a hummed tune or a bass line, and writes its notes onto a new MIDI track under it, lined up with the audio. It uses the pitch tracker behind the audio-to-MIDI page, and it has the same limit: one note at a time, so a full mix or a strummed chord has no single pitch to follow. New MIDI clip, in the Project menu or a track's menu, puts four empty bars at the playhead and opens the note editor on them.

Add a voice-over line with AI speech

Add speech…, in the Project menu or a track's menu, turns a typed line into a spoken clip at the playhead, with any of the 41 voices from the text to speech page (English, Spanish, French, Italian, Portuguese and Hindi) and a speed control. It lands on the selected track if that track is empty there, and otherwise on a new track, so it never cuts into audio you already have. The clip is ordinary audio from then on: trim it, fade it, put effects on it, duck the music under it. It is labelled as synthetic speech in its own file tags, and one undo takes it away. The voice model downloads once (326 MB for the fast GPU version, 92 MB for the smaller one) and is cached.

Process effects, tool pages and chord or key detection need audio, so a MIDI clip has Bounce to audio. It renders the clip through its instrument and swaps the notes for the recording, as one step of undo. What was measured, in a browser on every release: an imported C major scale exports every note within 2.4 cents of its pitch; every instrument plays A 440 at 440 Hz; the drum kit's kick sits under 200 Hz and its hi-hat over 5 kHz; a bounce sounds like the notes it replaced to within 5 cents and half a decibel; a harmonic melody converted to MIDI comes back as exactly its notes; and a project file keeps the notes in its project.json with no audio stored for them.

Notes are kept in seconds, not beats. Change the project tempo after adding MIDI and the notes stay where they were in time, rather than speeding up with the grid as they would in a DAW. Pitch bend, modulation and the other controllers in a file are skipped, and there is no velocity lane: notes keep the velocity they arrived with, and new ones get 96 out of 127.

Automation

Press A on a track to show its automation lane, then choose what it controls: volume, pan, a send, or any continuous setting of any effect on that track. Tap to add a point, drag to move one, and double-tap to delete it. For a fade under a voice, a filter that opens into the chorus, or a delay that throws on the last word, you can also turn on ● Auto in the mixer and simply move the control while the song plays. The movement is recorded, thinned to the fewest points that stay within 1% of what you did, and written into the lane. Automation belongs to the timeline, so when you cut a section out or insert silence, the lane moves with the audio after it.

What was measured

Every effect and every preset is rendered in a browser and checked on each release. The compressor takes 7.5 dB off a tone 10 dB over its threshold at 4:1, as the textbook says it should. The browser's built-in compressor, given the same settings, came out 9.3 dB louder than that because it adds make-up gain it does not report, which is why it is not used here. The limiter was pushed 18 dB into its ceiling and its true peak, measured the way the ITU-R BS.1770 loudness standard measures it, came out at the ceiling and not above. An EQ band at +6 dB boosts a tone at its centre by 6.00 dB. The multiband compressor's three bands sum back to within 0.15 dB of the input when they are not compressing. The limiter looks ahead 4.5 ms; every other track is delayed by the same amount so it stays in time, and the export trims that delay off the front. A mono track sits at −3 dB per side in the centre and at full level on one side when panned hard, the same pan law as before the mixer existed.

Process: effects that render

Process… keeps the one-off effects that need to see the whole file. They run the same code as the matching single-purpose page: loudness normalization (ITU-R BS.1770, checked against ffmpeg), noise reduction (spectral for steady hiss and hum, and an AI method that removes traffic, typing or barking behind a voice and leaves the clip exactly where it was), pause shortening, and speed and pitch changes through ffmpeg's time-stretch filter. Apply one to a whole clip, or select a range first and only that part changes. Speed and pitch need ffmpeg, which is a one-off 30 MB download the first time you use them.

Exporting

Export mixes every audible track together in 32-bit float, at the audio's own rate by default: a 96 kHz session exports at 96 kHz. You can pick any rate from 44.1 to 192 kHz instead. A clip at a different rate from the export is converted with the same windowed-sinc resampler as the sample rate converter, not the browser's own conversion. The formats are MP3 (the fast encoder, or LAME at V0 or 320 kbps), WAV at 16-bit, 24-bit or 32-bit float, FLAC at 16 or 24-bit, M4A and OGG. 16- and 24-bit files get triangular dither. An untouched 24-bit clip exported as 24-bit WAV comes out sample for sample the same as the file you put in; that is checked on every release. You can export the whole mix, the selected range, one clip, every track as a separate stem, or the timeline split at its markers, which is how you turn one long recording into numbered files. If the mix would clip, it is turned down so its loudest peak sits at -1 dB; if it would not, it is left exactly as you heard it. For a file going back into another editor, choose 32-bit WAV, which keeps levels above 0 dB intact instead of flattening them.

The export runs the same effect code as playback, offline and faster than real time. With Let reverb and echo ring out ticked, it keeps rendering after the last clip until the tails fall below −90 dB. A 2-second reverb on an 8-second clip ends up about 10.4 seconds long, instead of stopping dead. Stems include each track's own effects and its sends into the buses, and leave out the master chain unless you ask for it, so they can be mastered together somewhere else.

On a phone

Tap a clip to select it, then drag it to move it or drag its edges to trim. An unselected clip scrolls the timeline instead, because on a small screen one clip is often all you can see. Pinch to zoom, drag along the ruler to select a range, and press and hold a clip for its menu. The full-screen button gives the timeline the whole screen. Decoded audio is held in memory, so phones run out sooner than laptops; iPhones limit a tab's memory most tightly.

What it does not do

MIDI plays through the editor's own simple synth voices, not sampled instruments, and there are no third-party plugins (VST and Audio Unit plugins cannot run inside a browser) and no video output. There is one tempo and one time signature per project, so a song that speeds up or changes metre will drift off the grid. Two clips on the same track cannot overlap, so a crossfade is two tracks with a fade on each. Automation is drawn as straight lines between points, not curves. The real-time pitch shifter is built for harmonies and doubling; for a clean shift of a whole song, Process → Pitch uses ffmpeg offline and sounds better. A long chain on many tracks costs processing power, and an older phone will stutter before a laptop does. Turning off the effects you are not listening to helps. For surround, video or a full mastering suite, Reaper, Logic Pro, GarageBand or the Fairlight page in DaVinci Resolve are the next step up. They all open the WAV files this exports.