ISCA/ITG Workshop on Diversity in Large Speech and Language Models (LSLMs)
20 Feb 2025The workshop took place at Humboldt University of Berlin and set out to address topics such as:
- User diversity: Which aspects of human speech and language production affect the performance of LSLMs?
- Language use: How are LSLMs able to cope with different languages, dialects, and sociolects? How do they deal with code switching?
- Human adaptation: How does the use of LSLMs affect language comprehension and production? Which alignment effects occur, and in which time spans?
- Model adaptation: How do LSLMs need to be designed to cope with speech and language diversity? How do training and finetuning affect performance?
- Inclusion: What data and technologies are necessary to better cope with diversity in LSLMs?
With my colleague Erfan A. Shams, we presented our work on How Transformer-based ASR Models Handle Diverse Signal Perturbations at the Segmental Level. Together with Dorothea Kolossa (TU Berlin), Georg Rehm (DFKI), and Pierre Zweigenbaum (LISN), I participated in a panel discussion on future research directions that tackle diversity issues in LSLMs, which was moderated by Pia Knoeferle (HU Berlin) and Sebastian Möller (TU Berlin).
We concluded our trip with a concert at the Berliner Philharmonie. 🎶 Under the motto “Paradise lost?”, the orchestra and choir conducted by Marin Alsop intepreted pieces by Outi Tarkiainen, Brett Dean, Aaron Copland, and Heitor Villa-Lobos.