UniTalk

该数据集名为UniTalk，专为活跃说话者检测而设计，特别关注包含较少代表性语言、噪声背景以及拥挤场景等具有挑战性的真实世界情境。该数据集包含超过44.5小时的视频，并在48,693个说话者身份的帧级别提供了活跃说话者的标注。UniTalk为活跃说话者检测设立了新的基准，为在真实条件下开发和评估模型提供了宝贵的资源。该数据集规模超过44.5小时的视频，所涉及的任务是活跃说话者检测（Active Speaker Detection，简称Asd）。This dataset, named UniTalk, is specifically designed for active speaker detection, with a particular focus on challenging real-world scenarios including underrepresented languages, noisy backgrounds, and crowded environments. The dataset contains over 44.5 hours of video, and provides frame-level annotations of active speakers for 48,693 distinct speaker identities. UniTalk sets a new benchmark for active speaker detection, serving as a valuable resource for developing and evaluating models under real-world conditions. This dataset covers over 44.5 hours of video, and the target task is Active Speaker Detection (abbreviated as Asd).

搜集汇总

背景与挑战

背景概述

UniTalk-ASD是一个用于活动说话人检测的数据集，规模在100万到1000万条数据之间，包含视频、音频和CSV文件。数据集分为训练集和验证集，CSV文件提供视频ID、时间戳、人脸边界框坐标、说话状态标签（如SPEAKING_AUDIBLE或NOT_SPEAKING）等信息，适用于多模态说话人检测任务。