LLMs Outperform Embeddings in ASR Topic Matching Tasks

Lightweight LLM matchers using natural language descriptions outperform regex and sentence embedding encoders on unpunctuated, noisy ASR transcripts. The study evaluates topic-utterance judgments in contact center environments using both keyphrase and descriptive representations.