Data Mining Sound Archives: A New Scalable Algorithm for Parallel-Distributing Processing

This paper discusses a new algorithm, called the acoustic data-mining accelerator (ADA), which was developed to mine large sound archives for signals of interest including animal vocalizations. Background information on the development of ADA is provided, summarizing various projects that have utilized this technology since 2009. Performance was evaluated by comparing runtimes and efficiency metrics for two marine mammal detection algorithms that were applied to a 3-week single channel acoustic data set (sampled at 192 kHz and with 16 bit resolution). A total of four configurations (1, 8, 16 and 64 workers) demonstrated processing scalability. Results showed that each detection algorithm successfully processed the data set in all four configurations without changing the ADA algorithm. The fastest case (64 workers), had a total runtime performance of 1.5 hours; making the ADA 13 times more efficient than the serial case. Using a single worker it took more than 18 hours to process the same 3-week data set. Concurrent processing of both data-mining algorithms using 64 workers showed the highest efficiency gain (23x) compared to sequentially processing the data with a single worker.