Implementation of the Sound Classification Module on the Platform with Limited Resources

Abstract

There is a growing trend of using algorithms based on deep and machine learning in consumer devices, which imposes a challenge to the system's development because of the limited amount of resources in an embedded device. This paper presents integration of the sound classification module based on machine learning into a home audio system. The additional module enables dynamic change of processing controls according to the resulting confidence scores which indicate whether the current audio is speech, music or background noise. Main challenge of this paper is overcoming real-time processing constraints and embedded system's resource limitations. Results show that the sound classification module has been successfully integrated and produces the correct output.