DS³: Dual-Space Sample Selection for Speaker Representation Learning with Noisy Labels

📁 Datasets

Our experiments are conducted based on the VoxCeleb1 and VoxCeleb2 datasets.

Download VxoCeleb 1 & 2 datasets

Data lists

Download training lists with noisy labels

• The clean training list for VoxCeleb1 is vox1_clean.txt.
• The clean training list for VoxCeleb2 is vox2_clean.txt.
• The noisy training lists for VoxCeleb1 are formatted as vox1_[noisy_type]_[noisy_rate].txt.
• The noisy training lists for VoxCeleb2 are formatted as vox2_[noisy_type]_[noisy_rate].txt.
• The noisy training lists for VoxCeleb1K-O are formatted as vox1k_[noisy_type]_[noisy_rate].txt.
• The noisy training lists for VoxCeleb5K-O are formatted as vox5k_[noisy_type]_[noisy_rate].txt.
• The evaluation lists are vox_O.txt, vox_E.txt, and vox_H.txt.

📑 TODO

• Data preparation instructions released
• Release of core implementation
• Release of training and evaluation scripts