VoxCeleb2

VoxCeleb2 is a large-scale audio-visual dataset for speaker recognition, developed as an extension of VoxCeleb1. It collects speech from public YouTube videos and provides speaker labels across a diverse set of unseen environments and recording conditions. The dataset is designed to support robust speaker verification and identification research in real-world scenarios.

VoxCeleb2 contains thousands of speakers and hundreds of thousands of utterances, drawn from a large corpus of YouTube clips. The clips vary in language, accent, background noise, channel quality, and recording device. Each clip is labeled with the corresponding speaker identity, and many clips include video frames that can be used for auxiliary tasks such as face verification or lip synchrony, although the primary focus is on the auditory signal.

Standard experimental splits provide training, development, and test sets. The dataset is commonly used with state-of-the-art neural features such as x-vectors and deep speaker embeddings, evaluated using verification metrics like equal error rate and identification accuracy. Baseline results and open-source toolkits are widely cited in literature, making VoxCeleb2 a common benchmark for cross-domain and cross-language speaker recognition.

VoxCeleb2 was released by researchers from the Visual Geometry Group at the University of Oxford and collaborators. It is distributed for research use under terms that require citation and adherence to its licensing conditions. The dataset has been adopted by numerous studies in machine learning and speech processing to assess generalization to real-world data and cross-language variation.