High-Level Audio Features: Distributed Extraction and Similarity Search

Today, automatic extraction of high-level audio features suffers from two main scalability issues. First, the extraction algorithms are very demanding in terms of memory and computation resources. Second, copyright laws prevent the audio files to be shared among computers, limiting the use of existing distributed computation frameworks and reducing the transparency of the methods evaluation process. The iSound Music Warehouse (iSoundMW), presented in this paper, is a framework to collect and query high-level audio features. It performs the feature extraction in a two-step process that allows distributed computations while respecting copyright laws. Using public computers, the extraction can be performed on large scale music collections. However, to be truly valuable, data management tools to search among the extracted features are needed. The iSoundMW enables similarity search among the collected high-level features and demonstrates its flexibility and efficiency by using a weighted combination of high-level features and constraints while showing good search performance results.