Integration of GAN Environmental Sound Filtering and the CNN Sound Recognition System

With the development of deep learning in artificial intelligence, many finished products for voice recognition applications are available, such as Shazam recognizes music tunes, Apple Siri, Google Home, Amazon Alexa, and so on, which rely on human voice to execute actions. The embedded systems perform to customize services by using human voices when recognizing. However, their recognition may be affected by environmental sound. This study seeks to employ the conditional generative adversarial network (CGAN) to eliminate noise and residual network 18 (ResNet-18) to increase recognition accuracy by incurring the background sound and removing noise. We have developed the Nvidia Jetson AGX-Xavier framework to integrate an adverse network with the residual network connecting an omnidirectional microphone. The goal is to denoise the background sound and increase the recognition accuracy in a real- time environment for IoT applications.