Personalized Keyword Spotting at Microsoft Azure
Wake-word detection that knows who's speaking, and resists noise and adversarial audio.
Research internship with Microsoft Azure Cognitive Services (Speech), Summer 2021.
- Built a personalized keyword spotting system fusing RNN-T keyword detection with Res2Net speaker verification, cutting false-accept rate 16% while staying resilient to noisy and adversarial input.
- Benchmarked CNN and RNN-T models against FGSM/PGD attacks on Google Speech Commands and Cortana datasets, quantifying robustness gaps that informed production model selection.
Mentors: Jian Wu, Yong Zhao.