Cloud-based audio fingerprinting service

Audio Fingerprinting allows the identification of a query audio clip by matching the query audio fingerprints against a reference database. Traditionally, the matching process, which is CPU and memory intensive, is implemented either on a single computer (which is confined by CPU and memory limits for large databases), or on a computer cluster in a proprietary manner (which has limited flexibility in scaling the database). We have implemented audio fingerprinting prototype software that can run in a Cloud environment, specifically Hadoop/MapReduce. Because the MapReduce framework is designed for stream data processing instead of database query, we discuss how we address this challenge as well as other challenges such as appropriate data input format and partitioning. A performance evaluation of the software on a real dataset of ~8500 songs and real Hadoop clusters is presented to illustrate its efficacy, where a batch query of 1000 60sec clips can be completed in ~50sec in addition to ~30sec of database loading time with a 12-node cluster configuration.