Maike Züfle

This master's thesis investigates the handling of out-of-vocabulary (OOV) words in full-duplex language models. Full-duplex models, which are speech models that can listen and talk at the same time, have shown strong performance across many speech and dialogue tasks, yet their robustness to OOV words remains underexplored. The thesis will analyze how different full-duplex architectures represent and process OOV tokens and how these representations affect downstream task performance. Both intrinsic evaluations and task-based benchmarks will be used to systematically assess model behavior in the presence of unseen or rare vocabulary. Based on these findings, the work will explore methods to improve OOV handling.

Requirements:

• Strong programming and debugging skills
• Knowledge of Python and Pytorch
• Knowledge of machine learning

Related Work

• F-Actor: Controllable Conversational Behaviour in Full-Duplex Models (https://arxiv.org/abs/2601.11329)
• Personaplex: Voice and Role Control for Full Duplex Conversational Speech Models (https://research.nvidia.com/labs/adlr/files/personaplex/personaplex_preprint.pdf)
• FD-Bench: A Full-Duplex Benchmarking Pipeline Designed for Full Duplex Spoken Dialogue Systems (https://arxiv.org/abs/2507.19040)