Automatic Speech Recognition and Pronunciation Error Detection of Dutch Non-native Speech: cumulating speech resources in a pluricentric language

• Improving the performance of Automatic Speech Recognition (ASR) on learner's speech in a non-dominant variety pluricentric language by cumulating resources from different varieties same language. Through transfer learning approach, knowledge dominant can be transferred to and this benefits non-native recognition. Introducing plausible pronunciation errors native corpus based evaluate Pronunciation Error Detection (PED) algorithms. The shortage large-scale learners’ corpora precise manual annotations are two major challenges for automatic L2 recognition error detection speech, especially languages. In these cases, collecting annotating large (L2 learner) all is often unattainable. study, we investigated ways addressing problems through conventional Deep Neural Network (DNN) ASR-based Netherlandic Dutch Flemish resources. First, show that ASR baseline system improved combining datasets. Next, learned models trained data, learners' model further improved. order PED algorithms absence learner data with annotations, introduced simulate errors. For found results much better GOP classifier than one data. produced worse when were merged while ASR, lower WERs attained. Whether adding beneficial, thus seems depend specific task used for. We discuss results, compare them those related research suggest avenues future research.