A large scale audio fingerprinting system

Abstract

This paper presents a large scale audio fingerprinting system, which can be used to retrieve intended song from more than a million audio tracks. A time offset of several milliseconds could cause a number of audio fingerprints disappear. To decrease this effect, the query clip is converted to several copies with different start times. The audio fingerprints are mapped to hashes. Based on the statistics of the hashes on 25,000 audio tracks, it is found the frequency of occurrence does not follow the uniform distribution. A compact index structure is used to retain the quick access and none memory waste. Our experiments demonstrate the performance of this large scale audio fingerprinting system, and all of the query audio clips are captured from practical environments by various mobile devices. An iOS application is proposed based on this system and its index dataset contains more than 1 million songs. © Springer International Publishing Switzerland 2013.