Task Specific Sentence Embeddings for ASR Error Detection

International audience This paper presents a study on the modeling of automatic speechrecognition errors at the sentence level. We aim in this study tocompensate certain phenomena highlighted by the analysis ofoutputs generated by an ASR error detection system we pre-viously proposed. We investigated three different approaches,that are based respectively on the use of sentence embeddingsdedicated to ASR error detection task, on a probabilistic con-textual model, and on a bidirectional long short-term memory(BLSTM) architecture. An approach to build task-specific sen-tence embeddings is proposed and compared to the Doc2vecapproach. Experiments are performed on transcriptions gen-erated by the LIUM ASR system applied to the French ETAPEcorpus. They show that the proposed sentence embeddings ded-icated to ASR error detection achieve better results than genericsentence embeddings, and that the integration of task-specificembeddings in our system achieves better results than the prob-abilistic contextual model and BLSTM models.