I run SongUp AI, a web app that writes and sings full songs from a prompt, your own lyrics or a hummed tune. Every song we generate is stored with the language it was made in, and I had never looked at that column properly. So I counted.
This is a short write-up of what the data says, and of the small decisions that decide whether a number like this can be trusted.
The headline numbers
From 8 October to 6 November 2026, people made 10,725 songs in 75 languages. That is a 30-day window, so it is a snapshot and not a trend.
- 89% of songs were not in English. English is 10.7% of the total.
- Persian is the largest language at 24.2%.
- Arabic is second at 13.8%.
- English is third at 10.7%.
- Hindi is 5.1%, Bengali 3.3% and Urdu 3.0%.
The top of the list is not where an English-first product would start: English is third, behind Persian and Arabic.
How I counted
The idea is simple: every song row has a language code, so the count is a group-by. One grouped query over finished songs in the window, leaving out ready-made library copies and songs with no language; only totals are published.
Then the percentage of each language is its count divided by the total, rounded to one decimal place. The work is in the choices around the count.
Four things that change the answer
1. What counts as a song. I counted generated songs, not unique users. One person making twenty songs in Persian counts twenty times. If you want to know how many people use a language, that is a different question and needs a distinct count of accounts.
2. Which language the row says. The count is only as good as the language field stored with each song. A stored value can differ from what you actually hear in an edge case, so I would use a count like this for the big picture and not for anything that needs per-song accuracy.
3. The window. Thirty days is short. A holiday, a post going viral in one country or one outreach campaign can move a single language by several points. I would not read the gap between Hindi (5.1%) and Bengali (3.3%) as a ranking, but the gap between Persian and everything else is wide enough to be real.
4. Rounding. Percentages rounded to one decimal can add up to 99.9 or 100.1. If you publish a table, keep the raw counts next to the rounded percentages so nobody has to guess.
What surprised me
The most useful result was not the top of the list, it was the shape of it. A product built "for English first" would have one huge bar and a long tail. Ours has a small head of a few large languages (Persian, Arabic, English, Hindi) and then a long tail across 70 more. That affects real engineering decisions: where to spend time checking pronunciation, which scripts need right-to-left layout testing, and which languages deserve a dedicated landing page. Right-to-left rendering and Arabic dialect choice are not edge cases for us, they are the main case.
Try it or reuse the idea
If you run any product with user-generated content, add a language field early and count it early. It costs nothing and it will probably tell you something you did not expect about who your users are.
The figures and the date range are on one page: AI music by language. If you quote them, please include the window (8 October to 6 November 2026).
If you have run a similar count on your own product, I would like to hear which language surprised you.
Top comments (0)