OpenWER: Improving Cross-Lingual ASR Evaluation and Enabling Token-Based Accuracy Metrics

OpenWER is an open-source tool designed to enhance the robustness of Word Error Rate (WER) in cross-lingual Automatic Speech Recognition (ASR) evaluations. It introduces language-specific normalization and compound word detection, along with a token-based Levenshtein alignment that allows for more granular accuracy metrics, resulting in WER reductions of up to 25% across 52 languages compared to existing libraries. This advancement is significant for practitioners as it promotes fairer evaluations in ASR research, particularly for low-resource languages, thereby improving the reliability of multilingual models.