Automatic quality estimation for speech translation using joint ASR and MT features

Abstract

This paper addresses the automatic quality estimation of spoken language translation (SLT). This relatively new task is defined and formalized as a sequence-labeling problem where each word in the SLT hypothesis is tagged as good or bad according to a large feature set. We propose several word confidence estimators (WCE) based on our automatic evaluation of transcription (ASR) quality, translation (MT) quality, or both (combined ASR + MT). This research work is possible because we built a specific corpus, which contains 6.7k utterances comprising the quintuplet: ASR output, verbatim transcript, text translation, speech translation, and post-edition of the translation. The conclusion of our multiple experiments using joint ASR and MT features for WCE is that MT features remain the most influential while ASR features can bring interesting complementary information. In addition, the last part of the paper proposes to disentangle ASR errors and MT errors where each word in the SLT hypothesis is tagged as good, $$asr\_error$$ or $$mt\_error$$ . Robust quality estimators for SLT can be used for re-scoring speech translation graphs or for providing feedback to the user in interactive speech translation or computer-assisted speech-to-text scenarios.