HelloNPU: A Corpus for Small-Footprint Wake-Up Word Detection Research

As the first very step to activate speech interfaces, wake-up word detection aims to achieve a fully hand-free experience by detecting a specific word or phrase to activate the speech recognition and understanding modules. The task usually requires low-latency, highly accurate, small-footprint and easily migratory to power limited environment. In this paper, we describe the creation of HelloNPU, a publicly-available corpus that provides a common testbed to facilitate wake-up word detection research. We also introduce some baseline experimental results on this proposed corpus using the deep KWS approach. We hope the release of this corpus can trigger more studies on small-footprint wake-up word detection.