People rarely hum a tune perfectly when trying to identify a song. They whistle, or they sing with the wrong words. But according to a new arXiv preprint, the research community has largely ignored these everyday behaviors.

The paper argues that published query-by-humming systems and benchmarks are designed around hums and sung queries, leaving a gap between how people actually search and what systems are trained to handle. The authors note that no public dataset pairs whistles or misheard-lyric singing with the songs they are meant to retrieve.

That missing data matters for building robust music search tools. Without benchmarks that reflect real-world query variation, systems may perform well on clean hums but fail when users whistle a melody or mangle the lyrics. The preprint calls for expanding datasets to cover these common but overlooked query types.