What People Actually Make
With an AI Song Generator

Evasong offers two ways to start. You can describe the song in a sentence and let the AI write the lyrics, or switch to custom mode and paste words you already have. We expected the one-sentence mode to dominate. It did not.

Almost two thirds of users used their own lyrics at least once, and those lyrics were substantial: the median submission was about 1,300 characters, roughly the length of a full three-minute song. Half of all lyric submissions were already organized with section labels such as Verse and Chorus. These are people who have something specific to say and want it sung, not people looking for a random track.

Among the 664 people who described a song in their own words, these themes came up most often. One description can touch more than one theme.

The pattern is personal. Most requests are written for, or about, a specific person: a partner, a child, a friend with a birthday coming up. Weddings appear less often than we expected, probably because people describe them as love songs instead.

In custom mode, users can describe the sound they want. Of the 754 people who did, these were the most frequent requests.

Two things stand out. First, the voice matters more than the genre. Choosing who sings is the most common decision people make, and the split between male and female vocals is almost exactly even. Second, the most requested instruments are warm and acoustic: piano, acoustic guitar, violin. People are mostly making songs to give to someone, and those songs lean toward ballads rather than club tracks.

Instrumental music is a smaller but steady use case. 14% of users generated at least one track with no vocals, typically as background music for videos.

We looked at the language of each user's first song, using the lyrics or the description they wrote. Of the 1,192 users we could classify:

Six in ten users wrote in something other than English. Spanish is a clear second, and Arabic-script languages make up a group bigger than French. If you write in your own language, you are the majority, not an edge case. Modern AI vocals handle most of these languages well, and the result usually sounds more natural when the lyrics are written natively instead of translated.

Of the 412 users who turned their own lyrics into a finished song, 34% ran the same lyrics again. Almost all of them (91%) changed the style description before trying again.

In other words, people usually get the words right on their own and then use the AI to audition arrangements: a slower tempo, a female voice instead of a male one, piano instead of guitar. Treat the first result as a draft of the sound, not a final verdict on your lyrics.

The data covers every music generation request on Evasong from April 3 to October 11, 2026: 3,275 requests from 1,404 accounts. Each percentage counts people, not requests, so one very active user cannot skew a result. We only looked at aggregate counts. No individual lyrics, prompts, or songs are quoted or published here.

Topics and languages were classified by keyword matching, so treat those figures as close estimates rather than exact counts. Our generator fills in "Pop" when a user picks no style at all, so that default was excluded from the style analysis. Style figures come only from styles people typed or selected themselves.