Reliability of non-native speech automatic segmentation for prosodic feedback

This paper investigates the reliability of phonetic boundaries obtained through automatic segmentation of non-native speech for automatic prosodic feedback for foreign language learning. Indeed, prosodic feedback requires checking the fundamental frequency and the duration of phonetic segments of the learner utterances with respect to some reference patterns. Segmentation evaluations carried out on non-native speech data show that the automatic segmentation process takes benefit from the introduction of non-native pronunciation variants, and that several phonetic boundaries obtained through automatic segmentation seems to be reliable enough for providing relevant prosodic feedback. This concerns for example boundaries between obstruent sounds, such as plosives and fricatives, and vowel sounds.