Speaker: Wiliam Fernando López Gavilánez.
Abstract:
Audio classifiers designed for deployment across diverse devices often face unforeseen conditions during inference, attributable to device-specific characteristics. These challenges stem from variations in microphone transfer functions or on-chip digital signal pre-processing, which result in distribution shifts between training and inference data. In this talk, we review State-of-The-Art techniques for device-robustness employed in the DCASE Challenge for Data-Efficient Low-Complexity Acoustic Scene Classification. To further address these device-related issues within the Okey Aura Wake-up Word task, we introduce a novel device diversity database to evaluate and measure performance variations across different devices. Additionally, we used the parallel recordings from the TAU Urban Acoustic Scenes database to assess probability shifts between devices