The effect of sample rate reduction and audio compression on noise metric accuracy and statistical learning classifier performance

• Published in: The Journal of the Acoustical Society of America 2016-10, Vol.140 (4), p.3424-3424
• Main Authors: Blevins, Matthew G., Nykaza, Edward T., Netchaev, Anton
• Format: Article
• Language: English
• ISSN: 0001-4966; 1520-8524
• DOI: 10.1121/1.4971019

With the increasing availability of low-cost and portable noise monitors, the amount of captured data can reasonably be predicted to grow exponentially. The financial and computational expense of portable data storage, processing hardware, and wireless or cellular upload bandwidth, however, necessitate data compression techniques to ensure feasible continuous operation. Sample rate reduction causes a loss of high frequency components while compressed formats (such as mp3) remove parts of the signal that are inaudible to humans, but may be important to statistical learning algorithms. The degree to which these data reduction techniques degrade the accuracy of metrics calculated from the signals (e.g. peak level, equivalent level, sound exposure level, etc.) and the performance of statistical learning classifiers that use the metrics and/or band levels is examined. Over 20,000 recordings from noise monitors located on military installations are systematically down-sampled and compressed to use in the analysis. The root-mean-square error for each of over 40 noise metrics and the mean decrease in accuracy for a statistical learning classifier are presented.