Contribution
The (Not So) Universal Language: Linguistic Behavior in Text-to-Music Generation
* Presenting author
Abstract:
Henry Wadsworth Longfellow once wrote that “music is the universal language of mankind.”Modern text-to-music systems such as Suno and Udio appear to take that claim literally, often treating English as the language of universality.This pilot study examines how these models interpret geographic and linguistic cues in textual prompts.Using a controlled generation setup, songs are created for several world regions and musical genres under two conditions:(1) prompts naming only a region (“a song from [region]”) to reveal default linguistic behavior, and(2) prompts adding an explicit instruction (“… sung in [language]”) to test obedience.The resulting lyrics are automatically analyzed for detected language, code-switching, and English leakage.The study introduces a replicable framework for measuring linguistic bias in generative music models and situates language choice as an overlooked dimension of creative-AI fairness.By contrasting implicit and explicit language realization, it explores how “universal” AI music may in practice reproduce the dominance of a single global language, raising broader questions about cultural representation and accessibility in multimodal generation.