Exploring the Correlation of Pitch Accents and Semantic Slots for Spoken Language Understanding

Abstract

We investigate the correlation between pitch accents and semantic slots in human-machine speech. Using an automatic pitch accent detector on the ATIS corpus, we find that most words labelled with semantic slots also carry a pitch accent. Most of the pitch accented words that are not associated with a semantic label are still meaningful, pointing towards the speaker's intention. Our findings show that prosody constitutes a relevant and useful resource for spoken language understanding, especially considering the fact that our pitch accent detector does not require any kind of manual transcriptions during testing time.

End-to-End Spoken Language Understanding: Performance analyses of a voice command task in a low resource setting

“…We can confirm one of the conclusions of the study of Stehwien and Vu (2016) and Su and Tseng (2018) that prosodic information can point to the most important semantic information. We have shown that the E2E SLU model exploits prosodic information which favours its performance in predicting intents and in particular concepts.…”

“…Regarding speech, increasing and decreasing pitch contours help define prosody (Plack and Oxenham, 2005). According to Stehwien and Vu (2016) and Su and Tseng (2018), most of the words related to concepts also carry a pitch accent and can point to the most salient semantic information. In order to verify this, they studied the correlation between pitch variations and concepts.…”