Reproduce “Transsion Multilingual Speech Recognition System for MLC-SLM 2025 Challenge” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaoxiao Li, An Zhu, Youhai Jiang, Fengjie Zhu
Transsion's multilingual speech recognition system secured third place in the MLC-SLM 2025 Challenge, achieving a 9.83% Word/Character Error Rate across 11 languages. The system utilizes an Encoder-Adaptor-LLM architecture, combining a frozen Whisper-large-v3 encoder, a trainable adaptor, and a Qwen2.5-7B-Instruct Large Language Model fine-tuned with LoRA, demonstrating robust performance on conversational speech.
Key insights
- The Encoder-Adaptor-LLM architecture combining pre-trained speech encoders and LLMs, even with frozen base models, is highly effective for multilingual conversational ASR.
- Strategic data augmentation using external open-source datasets (e.g., MSR-86K) significantly improves model generalization and reduces error rates on challenging multilingual tasks.
- Parameter-efficient fine-tuning techniques like LoRA enable the successful adaptation of large LLMs for ASR tasks with manageable computational resources and without destabilizing pre-trained knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.