Efficient Acoustic Feature Extraction for Music Information Retrieval Using Programmable Gate Arrays

Many of the recent advances in music information retrieval from audio signals have been data-driven, i.e., resulting from the analysis of very large data sets. Widespread performance evaluations on common data sets, such as the annual MIREX events, have also been instrumental in advancing the field. These endeavors incur a large computational cost, and could potentially benefit greatly from more rapid calculation of acoustic features. Traditional, clusterbased solutions for large-scale feature extraction are expensive and space- and power-inefficient. Using the massively parallel architecture of the field programmable gate array (FPGA), it is possible to design an application specific chip rivaling the speed of a cluster for large-scale acoustic feature computation at lower cost. Recent advances in development tools, such as the Xilinx Blockset in Simulink, allow rapid prototyping, simulation, and implementation on actual hardware. Such devices also show potential for the implementation of MIR systems on embedded devices such as cell phones and PDAs where hardware acceleration would be an absolute necessity. We present a prototype library for acoustic feature calculation for implementation on Xilinx FPGA hardware. Furthermore, using a genre classification task we compare the performance of simulated hardware features to those computed using standard methods, demonstrating a nearly negligible drop in classification performance with the potential for large reductions in computation time.