From Bases to Exemplars, and From Separation to Understanding

Explore cutting-edge research in audio source separation and its applications in this comprehensive lecture by Paris Smaragdis from the University of Illinois at Urbana-Champaign. Delve into the evolution from basis functions to exemplars in audio processing, and discover how this shift enables direct computation of tasks like noisy speech recognition and music analysis without intermediate separation steps. Examine the geometric properties of mixed audio signals and learn about the use of large-scale decompositions with aggressive sparsity settings to achieve improved results. Gain insights from Smaragdis, a renowned expert in machine learning and signal processing for audio problems, as he presents his innovative work in this field. This 76-minute talk, delivered at the Center for Language & Speech Processing at Johns Hopkins University in 2012, offers a deep dive into advanced audio processing techniques and their practical applications.