yore-vocabulary-llm-filter

To improve Whisper speech-to-text accuracy for domain specific vocabulary, you can filter candidate terms using LLM classification to isolate technical jargon and project-specific names that require explicit vocabulary enrichment.

Whisper transcribes technical jargon and acronyms as common English words because its base vocabulary lacks domain specific context. Filtering phonetically ambiguous terms and adding them to a vocabulary file resolves these misinterpretations.

You can filter phonetically ambiguous terms by classifying harvested candidate words with LLM verdicts that distinguish terms needing retention from clear English or compound words suitable for dropping.

Yes, you can use an LLM to classify terms from a harvest file by assigning keep or drop verdicts, categories, and reasons to differentiate phonetically ambiguous technical terms from clear English words.

Yes, you need to manually review LLM vocabulary classification results before merging approved terms into a global vocabulary file to ensure accurate transcription of specialized domain terminology.

The best way to target project-specific names in speech-to-text transcription is to classify harvested terms using LLM filtering to isolate phonetically ambiguous identifiers and merge them into your Whisper vocabulary.