A text retrieval approach to content-based audio retrieval
Matthew Riley, Eric Heinen, Joydeep Ghosh · 2008
This paper presents a novel approach to robust, contentbased retrieval of digital music. We formulate the hashing and retrieval problems analogously to that of text retrieval and leverage established results for this unique application. Accordingly, songs are represented as a ”Bagof-Audio-Words” and similarity calculations follow directly from the well-known Vector Space model [12]. We evaluate our system on a 4000 song data set to demonstrate its practical applicability, and evaluation shows our technique to be robust to a variety of signal distortions. Most interestingly, the system is capable of matching studio recordings to live recordings of the same song with high accuracy. 1