Reproduce “Enhancing Large Language Model-based Speech Recognition by Contextualization for Rare and Ambiguous Words” 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
Kento Nozawa, Takashi Masuko, Toru Taniguchi
Preferred Elements, Inc. developed an LLM-based automatic speech recognition (ASR) system that improves accuracy for rare and ambiguous words by injecting keywords into text prompts. This method substantially reduced Keyword Error Rates (KWERs) for a 100-billion parameter LLM, dropping from 58.21% to 10.45% on the Japanese YODAS test set.
Key insights
- LLMs can effectively be prompted with keywords to provide contextual information for Automatic Speech Recognition (ASR), simplifying contextualization compared to traditional methods.
- Injecting keywords directly into the LLM's text prompt significantly enhances recognition of rare, ambiguous, and domain-specific words.
- Training data diversity and bias mitigation, such as handling the presence or absence of keywords during training, are essential for robust LLM-based ASR systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.