GPU-based acoustic feature extraction for electronic media processing

Multicore architectures are frequently utilized if very high computation power is required. At the same time current multicore graphic processing units (GPUs), designed for parallel data processing, have become applicable for general purpose computation. Thus, in current research projects the usage of GPUs is examined for a variety of applications. Thereby, GPUs are attractive for the realization of complex multimedia signal processing in terms of reducing computation time. An example for the processing of electronic media content is the automated classification of music collections, which is a highly attractive feature for multimedia terminals. Such tasks are based on the extraction of acoustic features, which are required to analyse the audio content. The extraction process is highly computation intensive and can benefit from the parallel computation power of GPUs. In this work, scalable parallelization methods are presented for GPU-based feature extraction applied to huge databases. The advantages of such a GPU realization are verified by a quantitative comparison to the results of a single core processor implementation in terms of computation times.