CONF Viglino_INTERSPEECH_2019/IDIAP End-to-End Accented Speech Recognition Viglino, Thibault Motlicek, Petr Cernak, Milos EXTERNAL https://publications.idiap.ch/attachments/papers/2022/Viglino_INTERSPEECH_2019.pdf PUBLIC https://publications.idiap.ch/index.php/publications/showcite/Viglino_Idiap-RR-04-2022 Related documents ISCA - International Conference on Speech and Language Processing, Interspeech Graz, Austria 2019 2140-2144 10.21437 doi Correct pronunciation is known to be the most difficult part to acquire for (native or non-native) language learners. The accented speech is thus more variable, and standard Automatic Speech Recognition (ASR) training approaches that rely on intermediate phone alignment might introduce errors during the ASR training. With end-to-end training we could alleviate this problem. In this work, we explore the use of multi-task training and accent embedding in the context of end-to-end ASR trained with the connectionist temporal classification loss. Comparing to the baseline developed using conventional ASR framework exploiting time-delay neural networks trained on accented English, we show significant relative improvement of about 25% in word error rate. Additional evaluation on unseen accent data yields relative improvements of of 31% and 2% for New Zealand English and Indian English, respectively. REPORT Viglino_Idiap-RR-04-2022/IDIAP End-to-end Accented Speech Recognition Viglino, Thibault Motlicek, Petr Cernak, Milos accent embedding Accented speech end-to-end multi-task speech recognition EXTERNAL https://publications.idiap.ch/attachments/reports/2019/Viglino_Idiap-RR-04-2022.pdf PUBLIC Idiap-RR-04-2022 2022 Idiap Rue Marconi 19, Martigny March 2022 Correct pronunciation is known to be the most difficult part to acquire for (native or non-native) language learners. The accented speech is thus more variable, and standard Automatic Speech Recognition (ASR) training approaches that rely on intermediate phone alignment might introduce errors during the ASR training. With end-to-end training we could alleviate this problem. In this work, we explore the use of multi-task training and accent embedding in the context of end-to-end ASR trained with the connectionist temporal classification loss. Comparing to the baseline developed using conventional ASR framework exploiting time-delay neural networks trained on accented English, we show significant relative improvement of about 25% in word error rate. Additional evaluation on unseen accent data yields relative improvements of of 31% and 2% for New Zealand English and Indian English, respectively.