A modular framework for efficient sound recognition using a smartphone

The identification of sounds is an important tool in ubiquitous and context aware applications. Today´s smartphones are capable of performing even computational intensive tasks, like digital signal processing and pattern recognition. In this contribution an implementation scheme and a framework for sound recognition for smartphones are presented. A basic sound recognition flow consists of preprocessing, feature extraction, feature selection, classiication, and action trigger. A flow is not hard coded but described in a JSON file and build dynamically by the framework. The framework itself is implemented in Java for the Android operating system. But specific algorithms can be realized in Java, C(++), and Renderscript for execution on the CPU, or in Filterscript for execution on a GPU. An example flow is presented and benchmark results are shown for Java-, C-, and Filterscript-implementations of Mel Frequency Cepstral Coefficients (MFCC). Recommendations for technology selection are made.