A Song-Making App Built for Persian

Most AI music tools are built for English, and it shows the moment you give them Persian: the words come back stiff, and the singer mispronounces half of them. This app was built the other way round. Persian is the language it is made for, and English, Spanish and the rest come along because nothing in the pipeline needs to be language-specific. You start from one of three places: your own words, a description of the sound you want, or the name of a poet whose voice you want the lyrics written in. If you bring no words, a lyricist agent writes them. A producer agent arranges them, a supervisor critiques the arrangement and sends it back when it falls short, and only then is the song rendered as audio. It runs as a phone app, used every day, with background and lock-screen playback, a shared radio of everyone's songs, playlists and offline downloads. The Persian work was measured, not guessed. Paid experiments were judged blind by a native speaker, and they showed that both the words and the voice were holding the songs back — and that telling the model to "enunciate clearly" made it worse. What shipped from that: the lyricist now writes to one held meter with proper rhyme, and before a Persian song is rendered the ambiguous words get their vowel marks, so the singer knows how to say them. Cost is visible at every step. Writing and arranging are free and can be redone as often as you like; only the final button spends money, about eight cents a song, and a daily cap stops anything running away. When the main music model is out of credit or not answering, a second one takes over — but never for Persian, which only goes to a singer someone has actually listened to.