Searching for Music Using Natural Language Queries and Relevance Feedback

We extend an approach to search inside large-scale music collections by enabling the user to give feedback on the retrieved music pieces. In the original approach, a search engine that can be queried through free-form natural language text is automatically built upon audio-based and Web-based similarity measures. Features for music pieces in the collection are derived automatically by retrieving relevant Web pages via Google queries and using the contents of these pages to construct term vectors. The additional use of information about acoustic similarity allows for reduction of the dimensionality of the vector space and characterization of audio pieces with no associated Web information. With the incorporation of relevance feedback, the retrieval of pieces can be adapted according to the preferences of the user and thus compensate for inadequately represented initial queries. The approach is evaluated on a collection comprising about 12,000 pieces by using semantic tags provided by Au...

• Type: Conference
• Year: 2007
• Where: AMR
• Authors: Peter Knees, Gerhard Widmer
• Keywords: AMR 2007, Large-scale Music Collections, Multimedia, Music Pieces, Web-based Similarity Measures