Analysis of the sensitivity of the End-Of-Turn Detection task to errors generated by the Automatic Speech Recognition process

An End-Of-Turn Detection Module (EOTD-M) is an essential component of automatic Spoken Dialogue Systems. The capability correctly detecting whether a user’s utterance has ended or not improves the accuracy in interpreting meaning message and decreases latency answer. Usually, dialogue systems, EOTD-M coupled with Automatic Speech Recognition (ASR-M) to transmit complete utterances Natural Language Understanding unit. Mistakes ASR-M transcription can have strong effect on performance EOTD-M. actual extent this depends particular combination errors sentence featurization techniques implemented as part In paper we investigate important relationship for based semantic information characteristics speakers (speech profiles). We introduce Simulator (ASR-SIM) that models different types mistakes well speech profiles. use simulator evaluate sensitivity Long Short-Term Memory network classifier trained EOTD techniques. Our experiments reveal ways which model influenced by errors. corroborate only ASR-SIM useful estimate customized noisy scenarios, but it also be used generate training datasets expected error rates real working conditions, leads better performance.