maikezu/noisy-fleurs

Noisy-FLEURS是一个从FLEURS数据集派生出来的衍生数据集，专为评估噪声稳健的多语种语音模型而创建。该数据集添加了两种类型的真实噪声（babel和ambient），以模拟具有挑战性的声学条件。babel噪声由多人同时说话的噪声组成，ambient噪声包括来自MUSAN语料库的环境声音和音乐样本的平衡混合。该数据集仅包含FLEURS的测试集，不包含训练或验证数据。

Noisy-FLEURS is a derivative of the FLEURS dataset, created specifically for evaluating noise-robust multilingual speech models. The dataset includes two types of realistic noise (babble and ambient) to simulate challenging acoustic conditions. Babble noise consists of overlapping human speech, while ambient noise is a balanced mix of environmental sounds and music samples from the MUSAN corpus. The dataset contains only the test set of FLEURS, without any training or validation data.