Content-based Information Retrieval in Streaming Music*

In this paper we examine searching by content in broadcasted streams of musical data, where the querier defines a set of preferred musical pieces and receives a list of broadcasting feeds that contain music similar to the preferred set. Streaming environments impose challenging requirements for content-based music information retrieval as memory limitations do not allow for buffering, data accumulation “on-the-fly” makes pre & post processing not a possibility while high response time is a necessity. To address these requirements we devise an incremental version of an award winning feature extraction and similarity process and propose a song boundary detection method in order to increase similarity accuracy and reduce costly feature extraction and similarity calculations. Extensive experimental results verify our claims and illustrate the superiority of the proposed method, over a baseline approach, as well as the suitability of the method for the streaming environment.