Language Family-Based Connector Sharing for LLM-Powered Speech Recognition
A recent research article presents an innovative approach for efficient multilingual automatic speech recognition (ASR) utilizing large language models (LLMs). It details a connector-sharing method that organizes languages according to their linguistic families, enabling a single lightweight connector to cater to an entire family, thereby minimizing parameters while preserving performance. The authors demonstrate the effectiveness of their approach through two multilingual LLMs and two real-world speech datasets. Results indicate that family-based connectors surpass per-language connectors in terms of parameter efficiency and generalization. Submitted to arXiv (identifier 2601.18899), the paper highlights the significance of typological knowledge in model architecture and proposes that linguistic classification can improve parameter sharing in speech AI, encouraging further exploration of linguistic features for this purpose.
Key facts
- The paper proposes connector sharing based on language family membership for LLM-based ASR.
- Prior work trained separate connectors per language, ignoring linguistic relatedness.
- One connector per language family reduces parameter count.
- The method improves generalization across domains.
- Evaluated on two multilingual LLMs.
- Evaluated on two real-world corpora: curated and crowd-sourced speech.
- The paper is available on arXiv as preprint 2601.18899.
- It was submitted to the Computation and Language (cs.CL) category.
Entities
—