ArabicSpeech/ArA-DF-2026
收藏资源简介:
ArA-DF-2026是一个用于二分类音频分类的阿拉伯语语音深度伪造检测数据集:标签0表示真实语音,标签1表示伪造语音。公开的训练和开发分割包含标签,所有公开测试分割无标签且需外部评分。所有发布的音频均为16 kHz单声道PCM音频,以无损FLAC格式打包在WebDataset TAR分片中。数据集包含多个分割:train(26,268行,有标签,文件名包含方言家族和性别标记)、dev(107,428行,有标签,文件名包含方言家族和性别标记)、test_track_1_clean(160,233行,无标签,文件名匿名化)、test_track_1_condition(160,233行,无标签,文件名匿名化)、test_track_1_chain(160,233行,无标签,文件名匿名化)和test_track_2(141,939行,无标签,文件名匿名化)。Track 1提供三种评估变体,用于不同公开评估条件下的方言泛化评估;Track 2评估在独立保留声学条件下的鲁棒性。提交格式为每个测试分割提交一个CSV文件,包含id和label列。加载元数据可使用datasets库的load_dataset函数,加载音频分片需指定WebDataset数据文件路径。
ArA-DF-2026 is an Arabic speech deepfake detection dataset for binary audio classification, where label 0 denotes authentic real speech and label 1 denotes forged speech. The publicly available training and development splits include labels, while all public test splits are unlabeled and require external scoring. All released audio files are 16 kHz mono PCM audio, packaged in WebDataset TAR shards in lossless FLAC format. The dataset comprises multiple splits: train (26,268 samples, labeled, with dialect family and gender markers included in filenames), dev (107,428 samples, labeled, with dialect family and gender markers included in filenames), test_track_1_clean (160,233 samples, unlabeled, filenames anonymized), test_track_1_condition (160,233 samples, unlabeled, filenames anonymized), test_track_1_chain (160,233 samples, unlabeled, filenames anonymized), and test_track_2 (141,939 samples, unlabeled, filenames anonymized). Track 1 offers three evaluation variants for dialect generalization assessment under different public evaluation conditions, while Track 2 evaluates model robustness under independently held-out acoustic conditions. The submission format requires a CSV file for each test split, containing the columns "id" and "label". Metadata can be loaded using the load_dataset function from the datasets library, and to load the audio shards, the file path of the WebDataset data needs to be specified.




