GE2E

  • Generalized End-to-end Loss for Speaker Verification
  • new loss function
  • training of Speaker Verification models more efficient
  • Unlike TE2E, the GE2E loss function updates the network in a way that emphasizes examples that are difficult to verify at each step of the training process
  • pushes the Embedding towards the [centroid] of the true speaker, and away from the [centroid](centroid] of the true speaker, and away from the [centroid.md) of the most similar different speaker
  • does not require an initial stage of example selection
  • MultiReader technique