Automatic and manual labeling of the corpus. All the SR and SH sentences were automatically segmented and phonetically annotated by our speech-text alignment tool based on acoustic Hidden Markov Models. A part of the aligned sentences were manually checked at phones and words levels (phonetic transcription) and corrections were made if necessary. The French sentences uttered by German and French speakers were corrected by seven French annotators (undergraduate students in phonetics), managed by a French expert phonetician (an assistant professor in phonetics). Annotators must add or remove a phone label and change a label when the speaker uttered a speech segment different from the canonical pronunciation. In case of voicing or devoicing of a consonant, annotators must add a diacritical mark. Moreover, they must carefully verify the phone boundaries and move them if necessary. When the boundary set by an annotator would be arbitrary he/she should use a diacritical symbol to mark it as fuzzy (as, for instance, the boundary between /a/ and /R/ in the word “départ”). Since the phonetic segmentation has been checked and corrected by seven annotators, it was necessary to verify the consistency of the seven annotators with the expert annotator.
