Dissertation Kristina Tesch

23 February 2024, by David Mosteller

Photo: Tesch

Kristina Tesch successfully defended her thesis Non-linear Spatial Filtering for Multi- channel Speech Enhancement and Separation supervised by Prof. Timo Gerkmann of the Signal Processing (SP) research group. Special thanks goes to the additional reviewers Chris Biemann and Reinhold Haeb-Umbach.

Abstract:

When speech signals are recorded with multiple microphones, not only temporal-spectral information but also spatial information can be used to reduce background noise and interferers. While traditional approaches combine a linear filter with a separate temporal-spectral postfilter, modern deep neural networks (DNNs) process spatial and temporal-spectral information together with a non-linear processing model. The dissertation of Kristina Tesch investigates the advantages of such a joint non-linear processing model from a statistical perspective and demonstrates a significant performance gain for non-Gaussian distributed noise signals. To realize this advantage in practical scenarios, DNN-based non-linear spatial filters are developed and combined with a control mechanism that allows to flexibly determine the extraction direction. Finally,  a real-time demonstrator of such a non-linear DNN-based filter for speech extraction has been developed (video).