SynVox2

SynVox2是由日本国立情报学研究所等机构创建的隐私友好型合成数据集，旨在解决VoxCeleb2数据集因隐私问题而无法访问的问题。该数据集包含5994个说话人的语音数据，通过语言鲁棒性强的正交Householder神经网络（OHNN）进行说话人匿名化处理，以保护隐私同时保持数据的实用性。SynVox2的创建过程涉及复杂的语音生成和后处理技术，旨在提高数据集的隐私保护、实用性和公平性。该数据集主要应用于自动说话人验证（ASV）领域，以解决现有数据集在隐私保护方面的不足。

SynVox2 is a privacy-preserving synthetic dataset created by institutions including the National Institute of Informatics (Japan) and others, aiming to address the inaccessibility of the VoxCeleb2 dataset due to privacy concerns. This dataset contains speech data from 5994 speakers, and adopts the linguistically robust Orthogonal Householder Neural Network (OHNN) for speaker anonymization to protect privacy while preserving data utility. The development of SynVox2 involves sophisticated speech generation and post-processing technologies, with the goal of enhancing the privacy protection, utility and fairness of the dataset. This dataset is primarily applied in the field of automatic speaker verification (ASV) to resolve the privacy protection deficiencies of existing datasets.