AraDiaWER: An Explainable Metric For Dialectical Arabic ASR

Goal of this study

Word Error Rate (WER) is a metric that compares the performance of Automatic Speech Recognition Systems (ASR) by comparing the reference transcript to the hypothesis, without any account for linguistic rules or semantic patterns. AraDiaWER combines semantic factors and error mitigation concepts from five SoTA studies (MR-WER, WERd, eWER, eWER2, and CODA for ASR) to introduce an explainable scoring approach for researchers.​

Hypothesis

Accounting for dialect-specific linguistic and semantic analyses between the ground truth and hypothesis of an ASR system, should yield a more explainable and improved WER measure for Dialectical Arabic (DA) speech recognition.​

Timeline

Primary Components

How it works