SEP-28K

SEP-28K是一个专门用于检测播客中口吃事件的数据集，由苹果公司创建。该数据集包含超过28,000个标记有五种事件类型的片段，包括阻塞、延长、声音重复、单词重复和插入语。音频来源于公共播客，主要由口吃者采访其他口吃者。数据集的创建过程涉及手动筛选播客，使用语音活动检测器提取3秒间隔的片段，并由至少三名非临床人员进行标注。SEP-28K的应用领域包括帮助言语病理学家跟踪个体的流畅性，以及改进对非典型言语模式的语音识别系统。

SEP-28K is a dataset dedicated to detecting stuttering events in podcasts, developed by Apple Inc. This dataset contains over 28,000 segments annotated with five types of stuttering-related events, including blocks, prolongations, sound repetitions, word repetitions, and interjections. The audio is sourced from public podcasts, which primarily feature interviews conducted by people who stutter with other people who stutter. The dataset creation process involves manually screening podcasts, extracting 3-second segments using voice activity detectors, and having the segments annotated by at least three non-clinical personnel. Applications of SEP-28K include assisting speech-language pathologists in tracking individuals' speech fluency, as well as improving speech recognition systems for atypical speech patterns.