HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systems
HALAS is a new dataset containing human-labeled examples of hallucinations from seven state-of-the-art Automatic Speech Recognition (ASR) models on real earnings call recordings. Researchers found that all tested models hallucinate, with errors showing strong overlap. The dataset serves as a rigorous benchmark to test and improve methods for detecting these errors in real speech.