We conduct research on technologies to accurately extract “who is speaking, where they are, and what they are saying” from audio recorded via microphones. By leveraging both signal processing theory and machine learning, we aim to develop methods that function effectively in real-world environments.
Dereverberation
Sound reflected off walls or ceilings (reverberation) impairs speech intelligibility and reduces the accuracy of automatic speech recognition. We research methods to remove reverberation from recorded audio, recovering the original speech signal.
Sound Source Localization
By analyzing factors such as the time-difference-of-arrival of sound across multiple microphones, we estimate the direction from which a sound originates. Our goal is to develop robust estimation methods that perform reliably even in environments with reverberation and background noise.
Microphone Selection
In scenarios where numerous microphones are distributed throughout a space, we address the challenge of selecting the optimal microphones to best capture a target sound. We focus on developing methods for effective selection while maintaining computational efficiency.
Speech Separation
This involves isolating the voices of individual speakers from recordings where multiple people are speaking simultaneously. We also utilize speech representations derived through self-supervised learning.