An FPGA based parallel architecture for music melody matching

Abstract

We propose an FPGA-based high performance parallel architecture for music retrieval through singing. The database consists of monophonic MIDI files which are modeled into strings, and the user sung query is modeled as a set of regular expressions (regexp), with consideration of possible key transpositions and tempo variations to tolerate imperfectly sung queries. An approximate regexp matching algorithm is developed to calculate the similarity between a regexp and a string, using edit distance as the metrics. The algorithm supports user sung queries starting anywhere in the database song, not necessarily from the beginning. Using the proposed formal models and algorithms, the similarity between the user sung query and each song in the database can be evaluated and the top-10 most similar results will be reported. We designed the approximate regexp matching algorithm in such way that all terms of the regexp can execute concurrently, which perfectly fits the massive parallelism provided by FPGA. The FPGA implemented melody matching engine (MME) is a parameterized modular architecture that can be reconfigured to implement different regexps by simply updating their parameter registers, and can therefore avoid the time-consuming code re-synthesis. MME also includes an on-board DDR2 memory to store the database, so that they can be read in to calculate edit distances locally on the board. This way, each MME forms a self-contained system and multiple MMEs can be clustered to increase parallel processing power, with virtually no overhead. MME is evaluated using the query corpus of ThinkIT with 355 sung files and database of 5563 MIDI files. It achieves a top-10 hit rate of 90.7% and a runtime of 19.4 seconds, averaging 54.6 milliseconds for a single query. MME achieves significant speedup over software-based systems while providing the same level of flexibility.