Reproduce “Class LM and word mapping for contextual biasing in End-to-End ASR” 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
Rongqing Huang, Ossama Abdel-hamid, Xinwei Li, Gunnar Evermann
Researchers at Apple enhanced end-to-end Automatic Speech Recognition (ASR) systems by integrating dynamic contextual information and improving rare word recognition, achieving a 70.6% relative Word Error Rate (WER) reduction on named entities without degrading performance on general vocabulary. The approach combines class language model training, contextual Finite-State Transducers (FSTs) with score normalization, and a pronunciation-based word mapping technique.
Key insights
- Training E2E models with class LM tags alone can improve both general and named entity-specific recognition by learning structural transitions.
- Score normalization is crucial to prevent "anti-biasing" effects, allowing strong contextual biasing for named entities without degrading accuracy on regular utterances.
- Leveraging phonetic similarity through word mapping can effectively address the challenge of recognizing rare or OOV named entities by transforming them into more common, pronouncible subword sequences.
- Explicit linguistic knowledge, like FSTs and pronunciation lexicons, can be harmoniously integrated into E2E ASR systems to overcome their inherent limitations in handling dynamic context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.