Reproduce “VoxCeleb2: Deep Speaker Recognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Joon Son Chung, Arsha Nagrani, Andrew Zisserman

This work introduces VoxCeleb2, a large-scale audio-visual dataset with over 1 million utterances from 6,000+ speakers in unconstrained "in the wild" conditions, and develops VGGVox, a deep convolutional neural network embedding system. Training with VoxCeleb2 enabled ResNet-50 models to achieve an Equal Error Rate (EER) of 3.95% on the original VoxCeleb1 test set, outperforming previous benchmarks for robust speaker recognition.

Key insights

  • Training deep neural networks on a substantially larger and more diverse dataset, like VoxCeleb2, leads to notable improvements in speaker recognition performance under noisy and unconstrained real-world conditions.
  • Deep ResNet architectures, when appropriately adapted for spectrogram inputs with temporal invariance, can effectively learn compact and robust speaker-discriminative embeddings.
  • A two-stage training approach with offline hard negative mining is effective for learning embeddings, especially for large datasets, by focusing on more challenging negative pairs to refine model discriminability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.