FormalASR: End-to-End Spoken Chinese to Formal Text

In three linesFormalASR introduces two compact models (0.6B and 1.7B parameters) that directly transcribe spoken Chinese into formal written text without an ASR+LLM pipeline. Fine-tuned on WenetSpeech-Formal and Speechio-Formal using supervised fine-tuning of Qwen3-ASR, they achieve 37.4% relative CER reduction over verbatim baselines and improve ROUGE-L and BERTScore.

Summary generated by Claude — human-verified