I’ve been making subliminals since 2018. More than 24,000 people have one of mine.
So when I explain how to make a subliminal, it comes from building thousands of them, and from listening to the home-made ones people send me asking why they don’t sound right.
Fair warning. The science on subliminal audio is thin. Some studies show small priming effects, plenty show nothing at all, and anybody selling you certainty is guessing. What I can give you is exactly what sits in the file and how to check with your own ears that you built it correctly. You decide what that’s worth.
New to this? Start with what subliminals actually are, then come back.
What goes into a subliminal
A subliminal is an ordinary audio file with a few things stacked inside it.
Affirmations. The script. Short present-tense statements, repeated for the length of the track. This is the actual content and it deserves most of your attention.
Voices. One or more readings of that script, in your voice or studio voices.
A masking bed. Music, rain, ocean, whatever you like. It sits over the voices and covers them, and the relationship between those two levels is the whole craft.
A silent layer (optional). The same voices riding on an ultrasonic tone around 18.5 kHz, above conscious hearing for most people. The classic “silent subliminal” technique.
Tones (optional). Two sine waves a few Hz apart, one per ear. On headphones you get a slow beating sensation. Or a single tone pulsed on and off, which works on speakers too. I’ll describe what they are and leave the claims about what they do to other people.
Five layers, and only the first three are required.
How to make a subliminal: the five steps
Step 1: Write the script
Everything downstream of a bad script is wasted effort, so start here and take your time.
Write in the present tense, first person. “I am calm in front of a crowd,” rather than “I will become calm.” You’re describing a state that already exists, so write it as a fact.
Write toward what you want. “I am not afraid of speaking” contains the word “afraid” and the image of fear, and the negation is the flimsiest part of the sentence. Say “I speak easily and my body stays relaxed.” Same goal, no ghost of the old thing hanging around in it. Same applies to lose, quit, stop, debt, pain, struggle. Rewrite around them.
Keep lines short. One idea per line. A line that runs three clauses long reads as a paragraph and loses its punch by the fortieth repeat.
How many lines? Between 50 and 250 is the sweet spot, and it took me years to settle on that range. Under 50 you hear the loop as a loop, even buried. Past 250 the script starts repeating itself in meaning and each line comes around so rarely that the track feels diffuse. I write 100 to 150 for most goals.
Here’s the maths that tempts people into huge scripts. Say your lines average 3 seconds. 100 lines is a 5-minute pass, so a 20-minute track plays every line 4 times. 500 lines is a 25-minute pass, and your 20-minute track never finishes the script once. So more lines buys you fewer repeats of each one. Repetition is the mechanism you’re betting on.
One production trick worth knowing: professional programs often carry the same idea in two perspectives, “I am confident” and “you are confident.” That doubles the spoken lines without doubling the concepts. Write everything in first person, then mirror it, and don’t count the mirrored copies when you’re deciding how long your script is.
Step 2: Give it voices
Now the script has to be spoken.
Your own voice carries something a synthetic voice can’t fake, and it costs nothing but time. It’s also the least forgiving option, because you’ll hear every breath and every line you rushed. More on this below.
Studio voices are faster and more consistent, and they let you do something you physically can’t do alone: layer several different speakers at once.
That layering matters more than people expect. One voice under music is a voice under music, and your ear can lock onto it and follow the words. Four voices reading at the same time turn into texture, because your attention can only track one talker and the rest become interference. It’s a well-studied effect, and it’s why a bigger cast can sit closer to the surface without becoming intelligible.
Spread them out. Give each voice its own spot in the stereo field, one left, one right, one just off centre. Two voices stacked in the middle fight for the same space and mostly just sound louder.
There are two sensible ways to hand a script to a cast. Split shares it out, so each line gets spoken once by one of the voices, which is the cheapest way to cover everything. Full has every voice read the whole thing. You get complete stacked layers, and you pay for however many voices you cast.
Repetition ends up the same either way. Eight voices sharing a 200-line script cycle through it as fast as one voice reading 25 lines. What changes is who speaks each line.
Step 3: Set the masking level
This is where most of the tracks people send me fall apart.
How far under the music should the voices sit? Get it wrong one way and you’ve made a strange-sounding audiobook. The other way, 20 minutes of rain.
Three things decide it.
One: your own ear is the first test. Play it back and listen normally. If you can follow the words without effort, the voices are too loud. If you can tell someone is speaking but can’t make out what, you’re in the right neighbourhood. If you’d never guess there was a voice in there, you’ve gone deep, which is fine if that’s what you wanted.
Two: more voices need less bed. A single voice wants roughly 15 dB of music above it before it stops being followable. Four voices want about 8 dB, because they’re already masking each other. Past four it barely moves, since the interference between voices has done most of the work. So the instinct to bury a big cast under a wall of music is backwards, and it’s why crowded home-made tracks so often sound like nothing but noise.
Three: the bed has to cover the right frequencies. Speech lives between roughly 200 Hz and 8 kHz, and the consonants that carry intelligibility sit around 1 to 4 kHz. A bed with deep bass and airy highs but a hole through the middle looks loud on a meter while hiding almost nothing, because the voices poke straight through the gap. Rain and running water cover that band beautifully. A sparse piano pad with long spaces between notes leaves a window open every few seconds.
In the Studio this stops being guesswork. It measures the gap between your voices and your bed and watches the speech band while the track plays. Then it suggests a level for your exact cast on your exact bed. One click applies it. Override it whenever you like, but you’ll be overriding a measurement.
Step 4: Add the silent layer, and then verify it
The silent subliminal technique is old and the principle is easy: take your voice signal and use it to modulate a tone sitting above the range most people can hear, around 18.5 kHz. The words are physically in the file, encoded on the carrier, and most listeners hear nothing.
Three honest notes about it.
“Inaudible” means inaudible to most people. Hearing at the top of the range drops off with age, and plenty of teenagers pick up 17 kHz as a faint whine. Adults over 40 usually can’t hear 16 kHz at all. Cheap speakers roll off up there too. So the layer is silent for most listeners on most gear, and I won’t tell you it’s silent for everyone.
Verify it. This is the part I care about most, because a silent layer is unfalsifiable by design. You can’t hear it, so you’re taking somebody’s word that your words are in there at all. The silent subliminal maker in the Studio has an audible check that shifts the carrier down into hearing range, so you can play it and hear your own script coming back at you. Then you switch it off and render. Any tool that asks you to believe in a layer you can’t inspect is asking for a lot.
The format will kill it if you let it. I’ve measured this. MP3 at 128 kbps removes the carrier completely, so the layer is simply gone from the file. At 320 kbps it survives in usable shape, and that’s the safest MP3 rate for the job. WAV keeps everything.
The mechanism is worth understanding. Low bitrates apply a lowpass filter, and the encoder also starves very quiet sidebands of bits because it assumes those frequencies won’t be missed. Both effects land on exactly the part carrying your words. (You’ll read online that MP3 cuts everything above 16 kHz, and I repeated it for a while myself. Then I measured it: 320 kbps passes well past 16 kHz.)
If you use a silent layer, export WAV. That’s the whole rule.
Step 5: Render and keep the file
Pick your length, render, save it somewhere you own. A finished subliminal is a plain audio file, so no app can revoke it and no account can lock you out.
| Export format | Silent layer | File size |
|---|---|---|
| WAV | Intact | Largest |
| MP3 at 320 kbps | Survives in usable shape | About a quarter of the WAV |
| MP3 at 128 kbps | Gone | Smallest |
WAV if you used a silent layer. MP3 at 320 kbps is fine for everything else.
How to make subliminals with your own voice
Your own voice is a legitimate choice, and for some people it’s the only one they want. A few things I’ve learned recording them.
Record in one sitting if you can. Energy drifts across sessions, and a track assembled from four different days has four different moods stitched into it.
Read slowly and evenly. You’ll want to perform. Don’t. A dramatic reading has peaks, and peaks poke through the bed at the exact moments you’d rather they didn’t.
Leave a beat of silence between lines. Trimming pauses later is easy, inventing them is not.
Record the script more than once and layer the takes. Two separate takes, panned apart, sound like two people. The same take copied twice just gets 6 dB louder, and it goes hollow the moment the copies drift out of alignment. The Studio treats every recording as its own take for that reason, and stacking them costs nothing.
Mixing your voice with studio voices works well too. Yours in the middle, synthetic voices spread around it, is a combination I use often.
How to make a subliminal in Audacity
Audacity can do all of this. It’s free, and I built subliminals in it for years. Here’s the actual workflow.
- Write your script in a plain text file, one line per affirmation.
- Get voice audio. Record on a new track with Shift+R, or import TTS files generated elsewhere. Audacity won’t produce voices for you.
- Mix your takes down with Tracks > Mix > Mix and Render, so you’re working with one voice bus.
- Loop it to length using Effect > Repeat, or copy and paste until you hit your target duration.
- Import your music with File > Import > Audio, and repeat it the same way.
- Set the levels. Select the voice track, then Effect > Volume and Compression > Amplify, and enter a negative dB figure. All the masking judgement above happens here, entirely by ear.
- Pan with the L/R slider in each track’s header.
- For a silent layer: duplicate the voice track, lowpass it hard with Effect > EQ and Filters > Low-Pass Filter at about 2800 Hz, then open Tools > Nyquist Prompt and run the expression below.
- Export with File > Export Audio as WAV.
(mult 0.5 (sum 1.0 *track*) (hzosc 18500))That multiplies your voice against an 18.5 kHz sine. The sum 1.0 keeps the carrier itself in the signal, which demodulates far more reliably on ordinary gear than a suppressed carrier does, and the 0.5 holds the peaks under full scale. Lowpass the voice first or the upper sideband runs into the ceiling of a 44100 Hz project. Set the project rate to 48000 for headroom, and check your work with Analyze > Plot Spectrum.
Now the honest part about what that costs you.
Every level gets set blind. Audacity can’t measure the gap between your voices and your bed, and it can’t warn you that a scooped-out bed is hiding less than it looks. You’ll get there by ear over several passes and never quite know if you landed it.
Looping is manual labour, and you repeat it for every layer. Six voices means six tracks, each panned and leveled by hand. Change one affirmation and you re-record it, line it back up, and redo the loop.
The silent layer gives you no feedback at all. You run the Nyquist expression, you see a bump in the spectrum plot, and you hope.
Audacity gives you total control and charges you time. A dedicated subliminal maker handles the assembly and the measurement, so your attention goes to the script, which is the part that decides whether the thing is any good.
How to listen once it’s built
Quiet. Set the music at a level you’d happily leave on in the background. The voices are already positioned relative to it, so turning everything up just makes it loud.
Headphones for some layers. Two-tone binaural material only works on headphones, since it depends on each ear getting a different signal. Pulsed single tones work on speakers. For a silent layer, headphones are safer, because plenty of small speakers barely reproduce 18.5 kHz.
Short and regular. Half an hour a day is easier to keep up than a six-hour session on a Sunday. Pick a slot you’ll actually keep: the commute, or the hour before sleep.
Commit to a number of days. 21, 30, 60, 90, whatever fits the goal. A defined run gives you something to evaluate at the end. The Studio’s player counts the days and the streak for you, which mostly stops you wondering whether you missed Tuesday.
Sleep listening is fine if you find it comfortable. Quieter than you’d have it awake, on a sleep timer, and skip earbuds you’ll roll onto.
The mistakes I see most
- Voices too loud. If you catch yourself reading along, the mix isn’t finished.
- A bed with a hole in it. Beautiful ambient pads that leave the speech band wide open.
- Negation in the script. Every “I don’t” and “I no longer” is a line worth rewriting.
- A 1,000-line script. Each line comes around once an hour. The whole thing feels vague.
- One voice duplicated. Two copies of one take add level and go hollow.
- MP3 with a silent layer. At 128 kbps the layer is gone, and the file gives you no hint.
- Trusting a silent layer you never checked. If you can’t verify it, you don’t know it’s there.
- Rebuilding from scratch every time. Save the project. You’ll want to change three lines next month.
- Switching programs every week. Give one script its full run before you judge it.
If you hit something that isn’t on that list, the Help Center is where I keep the longer answers.
Make your own subliminal without the assembly work
I built the Subliminal Studio to take the mechanical half of this off your hands. The AI writes a first draft from your topic, or you paste your own script, and either way you read and edit every line before a single word gets spoken. That’s the whole point. There’s no path by which the audio contains something you didn’t see.
From there you cast from 40+ studio voices and place them across the stereo field. Browse over 200 licensed music beds by mood, or bring your own, and let the Studio measure the masking for your exact cast. Silent layer with an audible check. Own-voice recording if you’d rather do the speaking, and tones if you want them. Render as WAV or MP3 as often as you like, and the file is yours to keep.
Credits pay for the AI writing and the voices. They never expire, and a line you’ve already paid for is never charged twice. (Handing a long render to our server costs 1 credit per track minute, worth knowing on a phone.)
It’s free to start, so you can make your own subliminal tonight: 400 welcome credits, no card needed, and a finished demo program already loaded so you can press play before deciding anything.
If you’d rather have it built for you, I still write custom subliminals by hand, one goal and one person at a time.
And there’s a shelf of finished premium programs refined over the past 8 years.
Whichever way you go, start with the script. Everything after it is engineering.
