Researchers have developed Dual-Form ASR (DF-ASR), a novel framework that integrates inverse text normalization (ITN) directly into the automatic speech recognition (ASR) process. Unlike traditional cascaded systems, DF-ASR uses paired spoken-form and written-form supervision, enhanced by an LLM-driven workflow and a sequence-level objective called ITN-MWER. This approach aims to improve the accuracy and readability of transcripts, especially for semantically dependent numeric expressions, by enabling normalization to be considered alongside acoustic-contextual modeling. Experiments on Chinese speech data demonstrated that DF-ASR outperforms existing open-source ASR-ITN systems and remains competitive with closed-source references, while also offering reliable control over transcript forms. AI
IMPACT This research could lead to more accurate and readable speech-to-text systems, particularly for complex numeric expressions.
RANK_REASON The cluster contains an academic paper detailing a new technical framework for speech recognition. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
