Cross-layer Models for Low-Resource Conversational ASR

Attend this plenary lecture from JSALT 2025 exploring cross-layer models for automatic speech recognition in conversational settings. Learn how Barbara Schuppler from TU Graz applies cross-layer optimization principles from communications engineering to access meaning across multiple levels of speech information. Discover her group's hybrid approach that combines data-driven and knowledge-based methods to integrate pronunciation and prosodic variation into ASR systems, particularly effective in low-resource environments. Examine how classical systems enhanced with linguistic knowledge can outperform transformer-based models for short, fragmented utterances in conversational speech. Explore applications beyond ASR including pathological speech analysis, dementia prediction, and assistive speech technologies. Gain insights into quantitative analyses of prosody and pronunciation variation in conversational speech, and understand how phonetic and linguistic knowledge can be integrated into speech technology for educational and healthcare applications. The lecture addresses the growing demand for accurate ASR in dialogue systems as they evolve from transactional tools to socially interactive agents, while providing unique perspectives on human speech processing through conversational data analysis.