[SV] VoxCeleb: a large-scale speaker identification dataset

화자인증 벤치마크 데이터셋으로 VoxCeleb1, 2가 많이 쓰이는데, 그 중 1에 대한 논문.

• characterstics:
  • purpose: spekaer identification, speaker verification
  • text-independent
  • large-scale
  • real-world dataset
• summary:
  • extracted from Youtube
  • 1251 celebrities (thus for SID, 1251-way classification task)
  • 55% male, 45% female
  • train set: 1211 spks, 340hrs
  • test set: 40 spk, 11hrs