The HNI Forum is an interdisciplinary lecture series dealing with current research topics relevant to the Heinz Nixdorf Institute. The edition on April 20, 2023, was organized and moderated by Prof. Häb-Umbach and focused on recent developments in the field of machine processing of speech and audio signals.
The speakers were Dr. Lukas Drude, Amazon R&D Aachen, and Janek Ebbers from the Department of Communications Engineering, who each gave exciting talks on very current research topics.
Lukas reported on modern speech recognition architectures based on the all-neural end-to-end approach. These have meanwhile replaced the hybrid approaches, which still use separate acoustic and language models, in many applications. However, end-to-end methods face new challenges, such as adapting to a changing context (new words, new acoustic environments, etc.). In his talk he showed how those can be addressed effectively.
Janek Ebbers delivered a presentation on machine learning methods for recognizing sounds or other acoustic events. In his talk, he presented learning methods that allow to train powerful recognizers from only non-annotated or weakly annotated training data.
Many thanks to Lukas and Janek for the entertaining and informative insights into their research work.