Say it with a song: writing music with AI, and when you may use it

A song is a good way to say something. A birthday invitation nobody deletes, a thank-you to a teacher, a farewell for someone leaving, a Valentine’s message that took more thought than a card. Until recently that meant knowing a musician. It no longer does.

This is the whole process, from the first note you write down to a finished track — and, at the end, the one rule that decides whether you are allowed to use what you made. Everything here was checked in August 2026.

Start with the story, not the tool

The quality of the lyrics depends almost entirely on what you bring. Before you open anything, answer four questions:

• Who is it for? How you would describe them to someone who has never met them.
• What is your history? How you met, what you have done together.
• What feeling should it carry? Gratitude, longing, joy, apology.
• Are there places or symbols? A market, a matatu route, a shared joke, a song you both know.

Write the answers out before you start. The difference between a song that could be for anyone and a song that could only be for this person is entirely in these details.

Give the tool enough to make the song personal — a first name, a place, a memory — and no more. What turns detail into risk is the combination that identifies a person: a full name with a date of birth, an ID number, an email address. And the story is not only yours; the person in the song has not agreed to be described to a company.

Write the lyrics with a chat model

Any of the major chat assistants will draft lyrics. In our own testing DeepSeek produced the most personal results, and it is free to use, but this is a matter of taste rather than a measurement — try two and keep the one you prefer.

Give it everything you wrote down, then ask clearly:

Write a warm pop song for my friend Amina for her birthday. We met at the market in Nairobi two years ago. She loves sunsets and her garden. The song should express gratitude and deep affection. One verse, a chorus, and a bridge. The chorus should be easy to remember.

Fix the draft yourself

The first draft is usually better than expected and never finished. Read it properly.

• Repair the rhymes that are almost right.
• Add what the model could not know — the detail only the two of you would recognise.
• Change the structure if you want a second verse or a longer bridge.
• Keep some repetition. Chat models over-repeat, and in a chorus that is exactly what makes it stick.

If a version is wrong, say what is wrong and ask for a rewrite. Vague feedback produces vague revisions.

Decide the style — and translate it

Before you generate anything, decide what the song should sound like: genre, tempo, instruments, whether the voice is male or female, how the verses differ from the chorus.

Here is the step most guides miss. Name the artist to the chat model, not to the music generator. Ask the chat model to turn “like Adele” into a description — slow acoustic ballad, piano and strings, female vocals, restrained verses and a fuller chorus — and paste that into the music tool. Most music generators reject artist names, song titles and lyrics as prompts outright, so the name would only get your generation refused. The description works better anyway, because it tells the model what to do rather than who to imitate.

Generate the music

Suno is the most capable for a complete song with vocals. The free plan gives 50 credits a day, which Suno puts at roughly ten songs daily, resetting every day rather than accumulating. Output is watermarked, and free-tier songs are licensed for personal, non-commercial use only — and subscribing later does not grant rights to songs made before.

ElevenLabs Music is the one trained on licensed catalogues, which makes it the cleaner option for anything with rights attached. Its consumer app offers up to seven songs a day free with no card — but that app is iOS only, so on an Android phone it is not an option. On the web the free allowance works out at roughly eleven minutes of music a month, without commercial rights; those start at around six dollars a month.

Lyria 3, inside the Gemini app, makes 30-second tracks from a text prompt or an uploaded photo, generates its own cover art, and marks every track with Google’s inaudible SynthID watermark. It is available wherever the Gemini app is, for users over 18. Longer three-minute tracks belong to the paid Gemini tiers, and sources disagree on exactly how much the free tier allows — check what your own account offers. For a short invitation clip, 30 seconds is often all you need.

Udio is frequently recommended and we are leaving it out on purpose: downloads of audio and stems are currently suspended across all its plans during its transition after the label agreements. You can make something there and not take it with you.

The line between a gift and a jingle

This is the part that catches people, and it has nothing to do with how the song sounds.

A song for a friend’s birthday, a wedding, a farewell or a thank-you is personal use. The free tiers cover it, and you are fine.

A song for a shop is not. A jingle for your business, background music under an advertisement, anything in a monetised video or on a channel that earns — that is commercial use, even if the shop is one room and the song is a favour. Free tiers forbid it, and no amount of good intention changes the licence.

The practical answer is cheaper than people expect. Commercial rights start at roughly six dollars a month on ElevenLabs and ten on Suno. Subscribe for the month in which you make the shop’s song, download it, and cancel. What matters is that the song is generated while you are subscribed — rights are not granted backwards.

Step by step with Suno

• Create a free account.
• Choose Create, then Custom Mode.
• Paste your lyrics, either whole or verse by verse.
• Paste the style description — for example: acoustic pop ballad, emotional, piano and strings, female vocals, slow tempo.
• Generate, and listen properly before judging.
• If it is wrong, change one thing in the prompt and generate again. Changing everything at once teaches you nothing.
• Export the track.

A few things that help

Describe, do not name. Genre, tempo, instruments, voice, mood. Every time.

Try more than one tool. The same lyrics sound noticeably different on each, and the free allowances exist so you can compare.

Work in small steps. Generate, listen, adjust one element, generate again.

Keep what worked. A style description that produced a good result is worth saving for the next occasion.

If your only computer is a phone, our article on AI tools you can actually use on a phone covers what runs well and what does not. For the wider picture, see our AI skills section.