wenet-e2e / wespeaker

Research and Production Oriented Speaker Verification, Recognition and Diarization Toolkit
Apache License 2.0
599 stars 104 forks source link

pre-trained models trained on only voxceleb1 #289

Closed sidhantls closed 3 months ago

sidhantls commented 4 months ago

I'm benchmarking speaker embedding models, for speaker verifications, that are trained and tested on the voxceleb1 dataset. I am referring to the pre-trained models list here but it looks like these models are trained on voxceleb2? if this is the case, are there any pre-trained models that have been trained on voxceleb1?

JiJiJiang commented 3 months ago
  1. Sorry that we do not train any models on the voxceleb1 dev set, for it only has 1211 speakers, which is not enough for training a well-performed speaker verification system. You can try to train on voxceleb1 dataset using our codes and have a comparison.
  2. We follow the protocol used in the VoxSRC challenge series. Training on the voxceleb2 dev dataset while testing on the whole voxceleb1 dataset, including vox1-O, vox1-E, and vox1-H, is widely-used for benchmarking. You can consider this setup.