Minje Kim, an assistant professor of intelligent systems engineering at the Luddy School of Informatics, Computing, and Engineering, has been honored with the National Science Foundation Faculty Early Career Development CAREER Award.
The CAREER award supports the early career-development activities of those teacher-scholars who most effectively integrate research and education within the context of the mission of their organization. Kim was honored for his project “Personalized Speech Enhancement: Test-Time Adaption Using No or Few Private Data.”
Kim’s project focuses on improving speech enhancement systems using artificial intelligence. Specifically, he is developing a method of training a personalized model of speech enhancement while requiring no or few data samples from the test-time users, making the system more efficient for possible use on small personal devices, such as hearing aids or “smart” devices. Creating a personalized model that only requires fine-tuning training from users will allow speech enhancement systems to be more user-friendly.
“The award means a lot to me, because it means what I have been doing could be important for both the research community and society,” Kim said. “I have worked on ‘adaptive’ machine learning models for the speech enhancement problem for years. It is an intriguing machine learning problem where there is no straightforward way to setup a learning objective.”
For example, if a smart speaker is shipped to a user with an accent who places the item in a kitchen, it’s a challenge for the adaptive models’ goal to adjust to the accent as well as typical kitchen noises such as blenders and dishwashers. Developing a method to allow models to adapt to environments after they are sent out to users is a critical step in smarter, more user-friendly devices, especially given the challenges faced when test-time users and their environments are unknown.
“Throughout the process of formulating the idea as a proposal and discussing it with the program directors and my peer researchers, I came to think more about why this test-time adaptation, or personalization, is important,” Kim said. “General-purpose machine learning models can underperform when used for people who are not well represented in the training dataset, a group that often underrepresented socially as well. Personalization can improve the model’s performance for them, improving fairness in AI. I would like to use this as an opportunity to think more about the privacy issues when researchers blindly collect personal data from the users to improve the AI model’s performance.”
Kim is director of the Signals and AI Group in Engineering at the Luddy School, and he joined the faculty at IU in 2016.
“The faculty at the Luddy School have a passion for innovation and are leaders in their respective fields,” said Dennis Groth, interim dean at the Luddy School. “Minje’s research is pushing the boundaries of speech enhancement, and he’s richly deserving of the NSF CAREER award.”