Task Specific Sentence Embeddings for ASR Error Detection

This paper presents a study on the modeling of automatic speech recognition errors at the sentence level. We aim in this study to compensate certain phenomena highlighted by the analysis of the outputs generated by the ASR error detection system we previously proposed. We investigated three different approaches, that are based respectively on the use of sentence embeddings dedicated to ASR error detection task, a probabilistic contextual model and a bidirectional long short term memory (BLSTM) architecture. An approach to build task-specific sentence embeddings is proposed and compared to the Doc2vec approach. Experiments are performed on transcriptions generated by the LIUM ASR system applied to the ETAPE corpus. They show that the proposed sentence embeddings dedicated to ASR error detection achieve better results than generic sentence embeddings and that the integration of task-specific sentence embeddings in our system achieves better results than the probabilistic contextual model and BLSTM models.