Answers are grounded in this paper and cite their sources.
Detecting errors in speech recognition is important because these errors negatively affect the accuracy of downstream tasks. To detect recognition errors, we focus on context, label sequence, and topics. Words in a sentence contain the context. Context information is useful in detecting errors from speech recognition because the errors must be outside the scope of the context. Furthermore, recognition error detection is a sequential problem, namely a sequential labeling task. Therefore, handling a label sequence is useful for error detection. In addition, a target speech contains topics, such as a talk about music, sports, and so on. A word that does not match the topic is considered an error. Therefore, the topic information is also useful. In this paper, we propose a method based on BERT, CRF, and LDA from the viewpoints BERT is used for context recognition. Conditional Random Fields (CRF) are used for the sequence labeling task. Latent Dirichlet Allocation (LDA) is to capture topic information. In this study, we compared several combinations of them: BERT only, BERT-CRF, and BERT-LDA-CRF models. We evaluate the models with the Corpus of Spontaneous Japanese (CSJ) corpus. The experimental result shows the effectiveness of the integration of CRF and LDA into BERT.
No takes yet. Share an insight, caveat, or question.
Yokoyama et al. (2023) studied this question.
Synapse has enriched one closely related paper. Consider it for comparative context: