Improving Query by Singing/Humming Systems over GPUs

This paper presents the use of GPUs for implementing a parallelized comparison engine in a query-by-singing/humming (QBSH) system, which takes a user´s singing or humming input and returns the most likely song from a database of about 13,000 song tracks. To speed up the comparison, we employ repeating pattern removal to retain only unique tunes in the database. Moreover, we explore different parallel schemes in GPU for achieving the best efficiency without sacrificing the retrieval accuracy. With an optimum speedup factor of 16, we have successfully implemented a QBSH system that is publicly available from the internet.