语言家族影响：跨语言边界评估LLM驱动ASR

Large Language Model (LLM)-powered Automatic Speech Recognition (ASR) systems achieve strong performance with limited resources by linking a frozen speech encoder to a pretrained LLM via a lightweight connector. Prior work trains a separate connector per language, overlooking linguistic relatedness. A novel and efficient connector-sharing strategy is proposed, based on linguistic family membership, enabling the use of a single connector per language family. This approach significantly reduces the number of connectors required for multilingual ASR systems, optimizing resource consumption and training efficiency by leveraging linguistic similarities. The strategy not only simplifies the complexity of multilingual ASR systems but also facilitates knowledge sharing across languages, particularly benefiting low-resource languages by sharing connector parameters with higher-resource languages within the same family. The core of this method lies in designing a connector architecture capable of generalizing across multiple languages within the same linguistic family, with its effectiveness and performance empirically validated across various language families. Experimental results demonstrate that this linguistic family-based connector-sharing strategy substantially reduces model parameters and training costs while maintaining or even improving ASR performance. This provides a new avenue for constructing more efficient and scalable multilingual ASR systems, showcasing the potential for cross-lingual transfer learning within a shared linguistic framework.