Generative Ambient Sound: The Complete Guide, With Sound You Can Play

I'm a designer who builds things in code, not a musician or a sound engineer. This is what I learned making one app, written down in one place. Each part has its own longer article if you want the detail, and I link them as we go.

• 1. The problem: why neither music nor silence works
• 2. What a good background sound looks like
• 3. What "generative" means
• 4. There are no recordings
• 5. The rule that makes it play: sound happens when things touch
• 6. Why it stays calm
• 7. How a plain tone gets a body
• 8. Making a place out of nothing
• 9. How I check whether a sound is any good
• 10. Where it runs, and what it costs
• 11. How to use it, and how to judge any background sound
• 12. What this guide does not claim

1. The problem: why neither music nor silence works

You sit down to work and the room is quiet. Within minutes something has caught your attention: the fridge, a neighbour, your own chair. So you put music on. For a song or two it helps. Then the chorus arrives, or a track you know too well, and you notice you have been listening instead of working.

I can't work in silence and I can't work to music. That is how I ended up reading about it, and then building my own background sound.

Three things in a sound pull attention away.

Words. Speech and lyrics are the worst case, because you don't choose to process them. A review of 65 studies found that they disrupt reading more than any other background sound. Another study found that music with lyrics hurt concentration and attention.

Something to follow. A beat, a tune heading for a resolution, a loop you have learned. None of it asks you to listen on purpose, and part of you keeps track anyway.

Sudden change. A door, a voice, a notification. A large review of noise and performance found that stop-start noise does somewhat more damage than steady noise.

Silence removes the first two problems and makes the third one worse. With nothing in the room, every small sound arrives at full strength, and each arrives without warning. Hearing works by contrast: in a café you don't notice a dropped spoon, and in a silent flat you hear the fridge switch on two rooms away.

The longer versions of this part are Why Music Doesn't Help You Focus, The Attention Cost of Lyrics, Is a Beat the Enemy of Concentration? and Silence Is Not Neutral.

2. What a good background sound looks like

If words, a thing to follow and sudden change are the problems, the answer is a sound with none of them. Steady, low, and with nothing that asks for you.

The test is not "which sound is most pleasant". It is "which sound asks the least of me". A sound can be lovely and still be a bad background, because loveliness is something you notice.

The usual answer is noise. It helps to know the colours.

Noise is every pitch at once, at random. The colour says which end is louder. White is equally loud everywhere and sounds like hiss. Pink falls away gently towards the top and sounds like steady rain. Brown falls away faster and sounds like a deep rumble. Violet is the reverse of brown, thin and sharp. White tends to be tiring because our ears are sharpest in the upper middle, so it comes out bright. Brown is named after the botanist Robert Brown and his random jitter of particles in water, not after a colour. I wrote the full version in White, Pink, Brown, Violet.

Noise has two limits. The research on it is mixed: one 2024 review found that white and pink noise helped task performance a little in young people with attention problems and made it slightly worse in those without. And recordings of rain, waves or noise are finite, so they loop, and a loop you have recognised is one more thing to follow. Two articles take that apart: Why Rain Sounds Work, and Why Recorded Rain Eventually Doesn't and Why Lo-Fi Beats Stop Working After the Third Hour.

That leaves the option I built: sound with one steady mood, changing slowly, that never repeats in detail. Which brings us to the word.

3. What "generative" means

Brian Eno made the word popular. He described generative music as music that is "ever-different and changing, and that is created by a system". Nothing is stored and played back. A set of rules runs, and sound comes out for as long as it runs. It is closer to an instrument that plays itself than to a music player.

The idea is older than computers, and the history is short enough to tell in one go:

• In 1917 Erik Satie coined the term "furniture music", meaning music that is part of a room in the way a chair is.
• In the 1970s Eno played two short melodies of different lengths through tape machines so that they layered over each other, and called the result "really automatic music". Later came Music for Airports, built by layering tape loops of different lengths, with the line that defined the style: ambient music "must be as ignorable as it is interesting".
• In the mid-1990s software did the job: a program called Koan made music from rules as it went, and Eno published an album made with it.
• In 2008 Eno and Peter Chilvers released a phone app, Bloom, where each tap becomes a tone that the app keeps going.
• Long before electricity, people built wind chimes, harps played by the wind and a dice game for composing dances.

The trick behind the tape loops is easy to hear. Loops of different lengths drift against each other, so the combination takes a very long time to line up again.

My app is a small example of the same idea in a browser tab. I'm not putting it next to Eno's work. I'm using an idea he and others made usable. The longer stories are A Short History of Generative Ambient Music, Brian Eno's Airport and Music You Don't Have to Listen To and Wind Chimes, Aeolian Harps, and Machines That Play Themselves.

4. There are no recordings

Ambient Sound Online contains no audio files. Every tone, texture and echo is calculated in your browser at the moment you hear it, using the Web Audio API that modern browsers have built in. The whole app is one page, and it uses the same amount of data whether you listen for a minute or all day.

I didn't know this was possible when I started. I assumed a sound app meant a folder of sound files.

Most ambient apps work like a vending machine: someone recorded a track once, and every press of play hands you the same one. This works more like a cook with a recipe. The browser has a few basic ingredients. An oscillator hums a plain note. A filter makes a sound brighter or duller. A gain control fades it in and out. Wire a handful of these together and the browser makes the sound fresh while you listen. Nothing is stored.

You can check it yourself. Start Ocean, then switch your phone to flight mode. The sound carries on, because there is nothing left to download. The honest limits are in Does It Work Offline?: once the page is open the sound is made on your device, but you can't open the page without a connection.

The main write-up is There Are No Audio Files.

5. The rule that makes it play: sound happens when things touch

Here is where a picture does the work of a score. Shapes drift slowly across the screen, each on its own path, and a sound starts when two of them overlap. There is no list of events. What you hear is what is happening on the screen.

Every shape moves by itself and doesn't know about the others. When two overlap, they sound. When they drift apart, they stop. There are two kinds of sound. Long ones fade in while a shape overlaps something and fade out when it leaves. Short ones fire once, at the moment of contact.

I wanted the simplest rule that would still surprise me. Most generated music has a plan underneath, even if the plan is random: something counts, and notes land on the count. Here the busy and quiet moments come from the picture. When shapes crowd into one corner there is a lot of sound, and when they spread out it thins to almost nothing. The part I like most is that you can see it coming. Two shapes drift towards each other and you know a sound is about to start.

Each kind of shape has one job.

The colour changes from world to world, but the job of each shape stays the same. The cost of this design is control: I can't make a particular sound happen at a particular moment. I can only change the odds, which is what the Objects slider does. More shapes means more meetings.

Two articles go deeper: Sound Only Happens When Things Touch and Shapes as Instruments. There was one practical problem too: shapes hovering at the edge of contact made the sound stutter on and off. The fix is in Hysteresis: How to Stop a Sound From Flickering.

6. Why it stays calm

Chance alone gives you noise. What keeps this pleasant are the parts that are fixed.

One scale for everything. All the sounds in a world share the same small set of notes, so any two that meet will fit. Space, the weightless world, uses only the notes of one wide, open chord, so nothing ever seems to arrive and nothing ever clashes. That is where the calm comes from. It started on a scale with no home note, which floated but felt uneasy over hours: Whole Tones Have No Home.

A limit on how many sounds play at once. Put enough shapes on screen and many are touching at any moment. Each sound is pleasant alone, and ten together are a grey blur. So only a few long sounds may play together, and only one drone. When little is sounding, each sound gets a longer tail. When a lot is sounding, tails are cut short. My instinct was to add more layers, and almost every time the sound got better it was because I took something away.

A scale that slowly drifts. A small set of notes keeps everything in harmony, but after an hour you have heard every note many times. So the whole scale slides up and down by a small amount over a few minutes. A note that has already started stays where it is, because a sliding pitch sounds like a tape machine running down. Only new notes pick up the new position. I can't hear it as movement, but a note from now is pitched a little differently from the same note a couple of minutes ago.

A clock made from a drone. There is no fixed tempo, because a tempo is a beat, and a beat is something to follow. Instead, while a drone is sounding it gives a slow, soft pulse, and short sounds lean towards that pulse so they land together. Take the drone away and the clock goes with it. Space has no pulse at all.

Soft starts and slow fades. Nothing jabs at you. More on this in the next part.

The detail is in Voice Limits: Why Fewer Sounds at Once Is a Feature, A Scale That Slowly Drifts and The Drone Is the Clock.

One thing "never repeats" does not mean: endless novelty. The set of sounds is fixed. After a while you will know the whale call and the sonar ping. What keeps changing is when they arrive and what they arrive next to. It means there is no loop point, because there is no stored sequence to come back round to.

7. How a plain tone gets a body

This is the part that took me longest, and the part I would most want to save a beginner from.

A sine wave is a hearing test. It is a single pitch with nothing else in it, the beep in a hearing test. A cello, a voice or rain on a window is never one pitch. Each is a main note with a stack of quieter notes around it, and that stack is what lets you tell a flute from a guitar playing the same note. An early version of my Space world was made of bare sine waves, held for a long time, fairly high up. I thought pure tones would sound clean and airy. They sounded like a whistle that wouldn't stop. Sine waves are fine as an ingredient, and they can't be the whole sound. See What a Sine Wave Sounds Like, and for the four classic wave shapes, The Four Basic Waveforms.

Deep is not the same as dark. When a sound was too sharp, I turned the top end down with a filter. That made it darker, like talking through a pillow, but not deeper. A filter can only take away what is already in the sound. A sound is deep when it has real energy at the low end, and the simplest way to put it there is a second, much quieter copy of the note one octave lower. You still hear one note, but it has weight. Too much is mud, so the layer should be quiet enough that you wouldn't notice it alone and would miss it when it was gone.

The full article is The Sub-Octave: Where "Deep" Comes From.

A start that doesn't click. A sound that goes from silence to full volume instantly makes a click, because that sharp edge is itself a sound. Bring the volume up over a few milliseconds and the click is gone, and it still sounds instant. Go much longer and the note starts to swell in, which is a different effect. The end of a note clicks too if you cut it off at full volume.

I found this the slow way. The short sounds in my Space world had a nasty tick at the front, and I spent a long time changing the sounds before I understood the problem was how they were switched on. See Attack Time.

A bell is plain tones, spaced untidily. Building a sound by adding plain tones together is called additive synthesis. Space the tones neatly and you get something smooth, like a soft organ. Push them slightly off the neat spacing and the same stack starts to sound like struck metal. It takes surprisingly little: a hair off is enough for a bell, and further off gives you a rougher, hollower singing bowl. I assumed the bells for the Forest world would have to be recordings, and they are small stacks of plain tones instead.

The longer version is Additive Synthesis: Building a Bell Out of Plain Tones. The companion ideas are Subtractive Synthesis, which carves a sound out of noise, Karplus–Strong, which makes a plucked string from a delay, and Why a Fifth Plus a Soft Onset Is Always a Singing Bowl.

8. Making a place out of nothing

A sound without a space around it is dry and close, as if it were happening inside your head. Two things fix that.

Position. Stereo is usually talked about as width, a big spread-out sound. It can do a more useful job: tell you where something is. In my app each sound sits where its shape sits on the screen, and moves smoothly when the shape moves. Shapes that are close to you are drawn bigger, and their sound is allowed further out to the sides. With your eyes shut you can still roughly follow the picture. It also stops several sounds from piling up into one lump in the middle. Use headphones for this one.

On a phone's single speaker, left and right fold into the middle. Nothing breaks. You lose the positions and keep the sound. More in Stereo as Information and Headphones vs Speakers for Ambient.

A room. The usual way to get reverb is to record a clap in a real hall and use the recording as a stamp. I have no hall and no files, so the app builds its own stamp: a short burst of noise that fades away the way an echo would. Each world gets a different one. That is why Ocean sounds submerged, and why Forest has a tail so long that one bell is still ringing when the next begins. Space has almost no room at all, because space has no walls, only a single long echo. See Reverb Without a Recording.

Three worlds, one engine. Ocean, Forest and Space share the same shapes, the same drifting and the same rule that sound starts on contact. What changes is the material, as in this summary:

I first thought of them as three separate instruments. They are closer to one instrument with three sets of strings. Space is the one world where I bent the engine, with no pulse, fewer sounds at once and forced gaps of quiet, because without those it filled up and stopped feeling empty. Details are in Three Worlds, Three Sets of Rules, and for the question of designing a place that doesn't exist, The Sound of a Place That Doesn't Exist and What Space Really Sounds Like.

9. How I check whether a sound is any good

For a long time my Space world sounded wrong: sharp, thin, tiring after a few minutes. Ocean, built with the same engine, was fine. I could hear that something was off, and I couldn't say what, which meant I couldn't fix it.

What helped was comparing instead of judging. I took the same few measurements of the world I liked and of the one I didn't, and looked at where they disagreed. There were four, and none needs any maths: how much of the sound sits in the low end, how fast it reaches full volume, how far its loudest instant sticks out above its average, and where it balances between low and high.

The numbers explained why, and that gave me something to change. They did not tell me when to stop. A sound can match a pleasant one on all four counts and still be dull, so after each change I still listened. I wouldn't call this measuring unpleasantness. It is measuring distance from something you already know is pleasant. The full account is Can You Measure Why a Sound Is Unpleasant?.

Two facts about hearing explain a lot of what went wrong.

Quiet changes what you hear. Hearing isn't even across pitch. We are most sensitive between about 2 and 5 kHz, and low sounds need more energy before we notice them at all. At low volume the bass fades first, then the sparkle, and what is left sounds thin. Background sound is played quietly, so it is heard in the conditions where bass is weakest. Judge a sound at the volume you will really use. See Sound at Low Volume.

Tiring is about where the energy sits. Thin, high, held tones with nothing underneath wear on me fastest, and so does anything that sits where hearing is sharpest. I can't give a number of minutes, and I don't trust anyone who does. "Listener fatigue" is a working term among audio people, not a recognised medical condition. If your ears ring or feel dull afterwards, it was too loud, and that one matters. See How Long Can You Listen Before It Becomes Noise.

10. Where it runs, and what it costs

Light to load. The app is small and opens quickly, with no framework and no outside code behind the sound: the engine uses only what the browser already has. After that first load there is nothing more to download. A small server part counts visits and handles payment, and these articles live separately.

One click to start. Browsers refuse to make sound until you do something, so every sound app needs a start button. Mine is that one click, and no trick gets round it. See The Autoplay Policy, Explained.

What synthesis costs. After the page loads there is nothing more to download, which matters on a phone plan. The cost moves to your device's processor and battery, which I looked at in Data, Processor, Battery. Phones are a special case: when the screen goes dark the browser can stop the sound, and the notes on that are in Why Your Phone Screen Goes Dark and the Sound Stops.

The honest part about data. No sound is sent anywhere, in either direction, and the page never uses your microphone. But the page does use cookies for anonymous analytics and keeps a simple count of visits, which world is playing, and how long people listen. If the connection drops, those fail quietly and the sound carries on.

11. How to use it, and how to judge any background sound

Ocean is free and starts in one click, with no account. Forest and Space unlock together with a single payment, and there is no subscription.

If you want to try it as a work sound, this is what I do:

• Start it quietly, just loud enough that you stop hearing the room. Not louder.
• Use it for one full work session before deciding, because the first five minutes tell you little.
• Use headphones if you can. Ocean carries its weight in the low end, which small speakers barely reproduce.
• If something about it nags, treat that as information about the sound. Try brown noise for a day and see which one you can leave on longer.

If you want to judge any background sound, mine included, the questions are the same ones from the start of this guide:

1. Are there words in it?
2. Is there anything to follow, such as a beat, a tune or a loop you have learned?
3. Is there anything sudden?
4. Does it still sit quietly in the background at the volume you will really use?
5. After a full session, do you want it off?

Where the answer is no to the first three and yes to the fourth, it is probably doing its job.

For particular situations, there are separate write-ups: Sound for Coding, Sound for Reading vs Writing vs Email, Open Office Survival, and, with the care these topics need, Focus Sound for ADHD, Can Generative Sound Help You Fall Asleep and Ambient Sound and Tinnitus.

12. What this guide does not claim

It does not claim a medical effect. Sound can make a room easier to work in. It does not treat anything, and nobody can promise you that it will help you focus, sleep or feel better. Where there is research, I've linked it and said what it found.

It does not claim a number. I can't tell you how many minutes before a sound wears on you, how loud is right, or which noise colour you should use. I can tell you what I noticed and why I think it happened.

It does not claim to be a music-theory or audio-engineering course. I'm a designer who built an app. Where I've simplified, for example by letting all the partials of a bell fade together, I've said so in the longer article.

It does not claim that this is the only good way to make background sound. It is mine, and it is built to ask as little of you as I could manage.

And one honest note on phones: I haven't promised that sound continues with the screen locked. Browsers differ on this, and I'd rather you test it on your own phone.

Sources

• Generative music, Wikipedia
• Furniture music, Wikipedia
• Ambient 1: Music for Airports, Wikipedia
• Bloom (software), Wikipedia
• Vasilev, Kirkby and Angele, 2018: review of 65 studies on background sound and reading
• Shih, Huang and Chiang, 2012: background music with and without lyrics, and attention
• Szalma and Hancock, 2011: review of noise and human performance
• Nigg and colleagues, 2024: review of white and pink noise and task performance
• Colors of noise, Wikipedia
• Additive synthesis, Wikipedia
• Inharmonicity, Wikipedia
• Whole-tone scale, Wikipedia
• Panning (audio), Wikipedia
• Crest factor, Wikipedia
• Spectral centroid, Wikipedia
• Equal-loudness contour, Wikipedia
• Listener fatigue, Wikipedia
• Autoplay policy in Chrome