Skip to main content Skip to main navigation


Domain Adversarial Training for German Accented Speech Recognition

Carlos Franzreb; Tim Polzehl
In: DAGA 2023 - 49. Jahrestagung für Akustik. Deutsche Jahrestagung für Akustik (DAGA-2023), 49. March 6-9, Hamburg, Germany, Pages 1413-1416, ISBN 978-3-939296-21-8, DEGA e.V, 2023.


Accented speech poses difficulties to automatic speech recognition (ASR) models because of its phonetic differences with the speech corpora the models were trained on. This issue can be addressed by performing domain adversarial training (DAT) with an accent classifier, where the ASR model is encouraged to discard accent information. In this study, we perform a comprehensive evaluation of this training method with the current state-of-the-art ASR model for German speech, the Conformer Transducer. We analyze the effect of important parameters, such as when to branch out the classifier or how to weight the two tasks, from which we draw several conclusions that can serve as a guideline for future experiments with this framework.


Weitere Links