A new arXiv paper introduces PronCFT, a method for contextual biasing of out-of-vocabulary (OOV) words in LLM-based automatic speech recognition. The authors note that these systems frequently substitute rare words with familiar ones that sound alike, a failure mode that undermines transcription accuracy for names, technical terms, and other uncommon vocabulary.
The abstract identifies two factors that contribute to this error. The first, described in the available text, is that the model has not learned how an unseen word's sound maps to its spelling. The second factor is mentioned but not detailed in the truncated abstract. PronCFT is presented as a fine-tuning approach that incorporates pronunciation context to address this gap.
Because the source is a single arXiv abstract, the article is limited to the authors' stated motivation and proposed method. No experimental results or comparisons are available in the provided text, so the effectiveness of PronCFT remains unverified from this source alone.