Tag
This paper evaluates demographic and accent biases in phoneme-based ASR systems, specifically WhisperIPA and ZIPA, using phoneme error rate and a new Soft PER metric, revealing persistent disparities across languages and groups.