Preserve to Expand: Source-Aware Adaptation and Model Merging for Multilingual LLM-Based ASR
arXiv cs.CL · 27 Oct 2026, 03:56 AEST
Expanding multilingual ASR systems to new languages poses challenges in preserving previously supported languages. The study evaluates various strategies for adapting connector-based Speech LLMs while maintaining source-language performance.
Key facts
- Expanding multilingual ASR systems to new languages is challenging when adaptation must preserve previously supported languages without replaying source-language training data.
- In single-target expansion, target-only fine-tuning and standard TIES merging learn new languages but cause severe source-language forgetting, while SSU-Wanda+TIES gives the best source-preserving trade-off.
- In multi-target expansion, update merging provides stable source retention, whereas recursive expansion can improve target adaptation under favorable orders but is highly order-sensitive.
Read aloud
Expanding multilingual ASR systems to new languages poses challenges in preserving previously supported languages. The study evaluates various strategies for adapting connector-based Speech L L M s while maintaining source-language performance. Expanding multilingual ASR systems to new languages is challenging when adaptation must preserve previously supported languages without replaying source-language training data. In single-target expansion, target-only fine-tuning and standard TIES merging learn new languages but cause severe source-language forgetting, while SSU-Wanda+TIES gives the best source-preserving trade-off. In multi-target expansion, update merging provides stable source retention, whereas recursive expansion can improve target adaptation under favorable orders but is highly order-sensitive.
Sources in this cluster
- Preserve to Expand: Source-Aware Adaptation and Model Merging for Multilingual LLM-Based ASR
arXiv cs.CL · 27 Oct 2026, 03:56 AEST