Towards a typology of ASR errors via syntax-prosody mapping

Fabian Santiago (1) , Camille Dutrey (2) , Martine Adda-Decker (1)

1 LPP - LPP - Laboratoire de Phonétique et Phonologie - UMR 7018
2 LNE - Laboratoire National de Métrologie et d'Essais [Trappes]

This study explores automatic speech recognition (ASR) errors from a syntax-prosody mapping perspective. Our contribution is threefold: we propose (i) an ASR error study according to French syntactic structures, (ii) a quantitative evaluation of the syntax-prosody mapping in reference transcriptions (iii) a qualitative analysis of syntax-prosody mapping violations in ASR transcriptions. Results show that some morphosyntactic and syntactic components are particularly prone to transcription errors such as proper names or verbal nuclei. In addition, we found that transcription errors in the ASR hypothesis may violate the syntactic-prosodic mapping rules. Such conflicting patterns may be used as clues to automatically detect ASR errors.

Mots clés: ASR errors; Syntax-Prosody mapping; French intonation

Domaines: Linguistique

Fabian Santiago, Camille Dutrey, Martine Adda-Decker. Towards a typology of ASR errors via syntax-prosody mapping. G. Adda, V. Barbu Mititelu, J. Mariani, D. Tufiș & I. Vasilescu. Errors by Humans and Machines in Multimedia, Multimodal and Multilingual Data Processing. Proceedings of ERRARE 2015, Editura Academiei Române, pp.175-192, 2015. ⟨hal-01707576⟩