The dangers of parsimony in query-by-humming applications

Query-by-humming systems attempt to address the needs of the non-expert user, for whom the most natural query format -- for the purposes of finding a tune, hook or melody of unknown providence -- is to sing it. While human listeners are quite tolerant of error in these queries, a music retrieval mechanism must explicitly model such errors in order to perform its task. We will present a unifying view of existing models, illuminating the assumptions underlying their respective designs, and demonstrating where such assumptions succeed and fail, through analysis and real-world experiments.