eval-ursa-2-enhanced-multimed-hard-20260408-1933
收藏资源简介:
该数据集用于评估Whisper模型'ursa-2-enhanced'在'Trelis/multimed-hard'数据集上的表现。数据集包含音频样本(如源数据集提供)、参考转录文本、模型预测文本、词错误率(WER)和字符错误率(CER)等字段。特别地,数据集还包含实体标注(如解剖学、生物标志物、条件、药物、组织和程序等)及对应的实体CER。整体实体CER为19.55%,不同类别的实体CER从0.00%到44.55%不等。该数据集适用于语音识别模型的性能评估,尤其是在医学和组织相关实体识别任务中。
This dataset is designed to evaluate the performance of the Whisper model 'ursa-2-enhanced' on the 'Trelis/multimed-hard' dataset. It contains fields including audio samples (as provided by the source dataset), reference transcriptions, model-generated transcriptions, Word Error Rate (WER), and Character Error Rate (CER). Notably, the dataset also includes entity annotations (such as anatomy, biomarkers, conditions, medications, tissues, and procedures) and their corresponding entity-specific CER values. The overall entity CER is 19.55%, with entity CER values across different categories ranging from 0.00% to 44.55%. This dataset is suitable for performance evaluation of speech recognition models, particularly in medical and tissue-related entity recognition tasks.




