Feature Extraction and Database Design for Music Software

Persistent storage and access of sound/music meta-data is an increasingly relevant topic to the developers of multimedia software. This paper focuses on the design of music signal analysis tools and database formats for modern applications. It is partly tutorial in nature, and partly a discussion of design issues. We begin with a high-level overview of the dimensions of music database (MDB) software, and then walk through the common g feature extraction techniques. A requirements analysis of several application categories will allow us to carefully determine which features might be most useful for them. This leads us to suggest concrete architectural and design criteria, and to close by introducing several of our recent implemented systems. The authors believe that much current MDB software suffers due to ad-hoc design of analysis systems and feature vectors, which often incorporate only low-level features and are not tuned for the application at hand. Our goal is to advance the state of the art of music meta-data extraction and database design by fostering a better engineering practice in the construction of high-level feature vectors and analysis engines for music software.