ai-sage/TimeGround-1M
收藏资源简介:
TimeGround-1M是一个合成英语音频数据集,专为时间感知语音理解设计,涵盖时间定位、时间描述和定时摘要任务。数据来源于YODAS2英语片段,从24k小时的源池中经过语言和静音比例过滤后,选择了14k小时的音频。通过基于LLM的验证、确定性有效性检查和分层采样,为三个时间相关任务生成了合成注释。最终数据按音频时长分组,包括0-10分钟、10-20分钟、20-60分钟和60-120分钟等桶。数据集提供多个配置:tl用于时间定位(输入事件描述,输出时间间隔),td用于时间描述(输入时间间隔,输出描述),sum用于定时摘要(输入完整音频,输出带时间戳的分段摘要),以及recordings用于包含嵌套注释的长形式录音。数据集支持流式加载,音频文件以WAV格式存储,可通过HuggingFace数据集库使用。
TimeGround-1M is a synthetic English audio dataset tailored for time-aware speech understanding, encompassing three core tasks: temporal localization, temporal description, and timed summarization. The dataset is sourced from YODAS2 English segments, where 14k hours of audio were selected from a 24k-hour source pool after filtering based on language and silence ratio criteria. Synthetic annotations for the three time-related tasks were generated via LLM-based validation, deterministic validity checks, and stratified sampling. The finalized dataset is grouped by audio duration into buckets including 0–10 minutes, 10–20 minutes, 20–60 minutes, and 60–120 minutes. The dataset offers multiple configurations: `tl` for temporal localization (input: event description, output: time interval), `td` for temporal description (input: time interval, output: description), `sum` for timed summarization (input: full audio, output: timestamped segmented summary), and `recordings` for long-form recordings with nested annotations. Audio files are stored in WAV format, and the dataset supports streaming loading, which can be accessed via the Hugging Face Datasets library.





