DIY: make your own track or jingle with AI — for a video, an app, a podcast or a product — with rights you can actually use

Your video needs an intro. Your app needs a loop. Your podcast needs a sting, your brand needs a sound, your product page needs thirty seconds of something that isn’t a stock library everyone recognises. The reflex is to license a track or hire a composer. This is the recipe I run for a whole label, cut down to one track you can make tonight.

You end up with

A finished track or jingle in the length and mood you need, exported as audio, with the commercial rights to use it in your own work

What it costs

A paid plan on an AI music generator (low tens of euros a month). Marginal cost per track: well under a euro. The free tier does not give you commercial rights

How long

An hour for a jingle; an evening for a track with vocals you're proud of

Runs on

An AI music generator on a paid plan · a text editor for the lyrics · optionally a free audio editor for trimming and fades

Step 0 — pick the plan for the rights, not the features

Every generator has a free tier. On the major ones the free tier is non-commercial: you can listen, you cannot put the track under an ad, in an app or in a client’s video. Commercial use comes with the paid plan. That is the whole reason to pay; the higher quality is a bonus. Read the terms on the day you sign up, and screenshot them — terms change (mine changed the download cap mid-year).

Step 1 — write the prompt like a producer, not a fan

The model responds to production vocabulary, not adjectives. “Epic emotional track” gives you mush. This gives you a track:

phonk wave trap, dark, 145 BPM, sub-bass slides, distorted 808s, chopped vocal stabs, reversed bell plucks, detuned synth lead, half-time bounce, tape saturation, cold nocturnal mood

That is a real seed prompt from the label. The structure: genre + sub-genre, tempo, three to six named instruments or techniques, one mood phrase. If you don’t know the vocabulary, ask a chat assistant to describe a reference track you like in production terms and start from that — you’re not copying the track, you’re borrowing the language.

For a jingle add the constraints: “instrumental, 15 seconds, clean ending, no fade”. For an app loop: “seamless loop, no intro, no build”.

Step 2 — lyrics, if you want a voice

Write them yourself. The model will offer to; the result is the same four rhymes everyone gets, and it is the one place your brand actually gets to say something. Structure them with section tags the generator understands — [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Outro] — and keep lines short: eight to ten syllables sing; sixteen get mangled.

Two things I learned the hard way: numbers and names get sung the way they’re spelled, so write “eighteen fifty-six”, not “1856”; and a chorus that repeats exactly is what makes a hook — resist improving it in the second pass.

Step 3 — generate in pairs, keep the seed, pick with your ears

Generators return two variants per run. Run the same prompt three times before changing anything — six takes of one idea beats one take of six ideas. Pick with the sound off first (does the structure work at 0:00, 0:15, 0:45?), then with the sound on. If a take is 80 % right, most tools let you extend or remix that take rather than starting over.

Step 4 — export and finish

Export the highest quality the plan offers (WAV if available, otherwise the best MP3). In a free audio editor: trim silence, add a 200 ms fade at the end for a jingle, normalise loudness if it’s going next to speech. For a loop, cut on a beat and test the seam by playing it three times in a row.

Name the file with the version and the date. The label’s archive rule: every take you keep is a file and the text that made it.

Where it’s allowed to go — and where it isn’t

• Yes: your videos, your app, your podcast, your ads, your product page, your client’s project (with disclosure), your games — the tracks from the label became the soundtracks in my games for exactly zero extra cost.
• Carefully: YouTube. A track under your own video is fine; a channel that is nothing but generated music with no human layer is what the platform’s spam rules target.
• No, or not like this: dumping generated tracks onto streaming by the hundred. That is the “AI slop” the distributors and platforms are purging, and the per-stream payout wouldn’t have paid for the subscription anyway. If you want a catalog on streaming, that is a different project with its own recipe — and its own honest economics.

Where it breaks

• It sounds like everything else. The prompt was adjectives. Go back to Step 1 and name instruments.
• The vocal mangles a line. The line is too long or has a number in it.
• The “loop” has a click. You cut off-beat. Cut on the downbeat and cross-fade 10 ms.
• Your client’s platform flagged it. Some libraries and broadcasters refuse AI-generated audio; ask before delivering, and keep the generation history as your provenance.

Related: how to release music alone · the label log · all recipes in Do it yourself.