Reproduce “Dual-Form ASR: Semantics-Aware Inverse Text Normalization for Chinese Speech Recognition” 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
Fengrun Zhang, Li Fu, Wangjin Zhou, Lu Fan, Youzheng Wu, Xiaodong He
The Dual-Form ASR (DF-ASR) framework integrates semantics-aware inverse text normalization directly into the speech recognition process for Chinese, enabling prompt-conditioned generation of either spoken-form or written-form transcripts. This approach improves accuracy in required normalizations and better preserves forbidden spoken spans compared to traditional cascaded systems.
Key insights
- Integrating semantics-aware ITN directly into the ASR process, with joint optimization, significantly enhances performance and context-sensitivity compared to decoupled, cascaded approaches.
- Large Language Models can be effectively utilized to generate and verify high-quality dual-form supervision data at scale, overcoming the limitations of manual annotation and context-insensitive rule-based methods.
- A tailored sequence-level objective, such as ITN-MWER, is crucial for optimizing ASR models to manage the unique semantic complexities and error costs associated with numeric expressions during text normalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.