Reproduce “The Role of Prosody in Spoken Question Answering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jie Chi, Maureen de Seyssel, Natalie Schluter
Researchers investigated the individual and combined roles of prosodic and lexical information in spoken question answering (SQA) by creating controlled audio manipulations on the SLUE-SQA-5 dataset. Their findings indicate that while prosodic cues alone can guide answer span identification, current self-supervised models primarily leverage lexical information, with mid-frequency audio components (above 300Hz but below 1800Hz) also contributing.
Key insights
- Prosodic information alone contains meaningful cues for spoken question answering, enabling models to achieve reasonable performance in identifying answer spans.
- Current self-supervised learning models exhibit a strong bias towards lexical information, showing quicker convergence and lower loss when word content is available.
- Mid-frequency audio components (between 300Hz and 1800Hz) also contribute significantly to SQA performance, suggesting cues beyond pure prosody or full lexical content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.