Distributed Audio Feature Extraction for Music

One of the important challenges facing music information retrieval (MIR) of audio signals is scaling analysis algorithms to large collections. Typically, analysis of audio signals utilizes sophisticated signal processing and machine learning techniques that require significant computational resources. Therefore, audio MIR is an area were computational resources are a significant bottleneck. For example, the number of pieces utilized in the majority of existing work in audio MIR is at most a few thousand files. Computing audio features over thousands files can sometimes take days of processing. In this paper, we describe how Marsyas-0.2, a free software framework for audio analysis and synthesis can be used to rapidly implement efficient distributed audio analysis algorithms. The framework is based on a dataflow architecture which facilitates partitioning of audio computations over multiple computers. Experimental results demonstrating the effectiveness of the proposed approach are presented.