NbAiLab/nb-asr-eval-withwav-sorted
收藏资源简介:
该数据集名为“NB-ASR Eval with WAV Sorted”,是“NbAiLab/nb-asr-eval-withwav”的副本,专门用于针对人工清理,并按照自动语音识别(ASR)的词错误率(WER)难度进行排序。数据集主要用于挪威语(包括博克马尔语nb和尼诺斯克语nn)的ASR评估和测试任务。它包含多个分割(split),如测试集和验证集,每个分割都有对应的WAV音频文件和JSONL元数据文件。音频文件为单声道16位WAV格式,来源于共享存储。数据集结构符合Hugging Face AudioFolder格式,每个样本包含音频、唯一ID和参考文本。README还详细说明了数据集的过滤策略(例如,基于WER阈值过滤样本)、各分割的样本数量统计,以及使用示例。数据集旨在支持ASR模型的评估,强调测试集仅用于最终决策,不应在开发过程中优化。
The dataset is named NB-ASR Eval with WAV Sorted, which is a copy of NbAiLab/nb-asr-eval-withwav intended for targeted human cleanup, with rows sorted by ASR/WER difficulty. It is designed for evaluation and test splits in the nb-asr project, focusing on automatic speech recognition (ASR) for Norwegian languages (Bokmål nb and Nynorsk nn). The dataset follows a Hugging Face AudioFolder structure, with each split containing a metadata.jsonl file and referenced WAV audio files. Audio files are 16-bit mono WAVs materialized from shared storage. Each example includes fields: audio (Audio object), id (unique utterance ID), and text (reference transcript). The README details filtering strategies based on WER thresholds, sample counts for validation and test splits, and usage instructions. It emphasizes that test splits should only be used for final release decisions and not optimized during development.




