README

This repo contains Keras code used to train models for the task of speaker verification, as outlined in the paper below from ICASSP 2019:

[Utterance-level Aggregation For Speaker Recognition In The Wild (Xie et al., ICASSP 2019)].

Dependencies

• Python 2.7.15
• Keras 2.2.4
• Tensorflow 1.8.0

Data

The datasets used to train these models are the VoxCeleb datasets, which can be found at the link below.

• Voxceleb1, Voxceleb2

Training the model

To train the model on the Voxceleb2 dataset, please run

- python src/main.py --net resnet34s --batch_size 160 --gpu 2,3 --lr 0.001 --optimizer adam --epochs 48 --multiprocess 8 --loss softmax --data_path ../path_to_voxceleb2

Model

• All models are available at the following google drive link: https://drive.google.com/open?id=1M_SXoW1ceKm3LghItY2ENKKUn3cWYfZm
• Download the models and put them in the folder, model/

Testing the model

To test a specific model on the voxceleb1 dataset eg. the ResNet34 model trained using adam with a softmax loss, and feature dimension 512 please run

- python src/predict.py --gpu 1 --net resnet34s --ghost_cluster 2 --vlad_cluster 8 --loss softmax --ohem 2 --resume ../model/gvlad_softmax/2019-01-11_resnet34_bs142_adam_lr0.001_vlad8_ghost2_bdim512_ohemlevel2/weights-47-0.866.h5 

Citation

@InProceedings{Xie19,
  author       = "W. Xie, A. Nagrani, J. S. Chung, A. Zisserman ",
  title        = "Utterance-level Aggregation For Speaker Recognition In The Wild.",
  booktitle    = "ICASSP, 2019",
  year         = "2019",
}

@InProceedings{Chung18,
  author       = "J. S. Chung*, A. Nagrani*, A. Zisserman ",
  title        = "VoxCeleb2: Deep Speaker Recognition.",
  booktitle    = "INTERSPEECH, 2018",
  year         = "2018",
}

@InProceedings{Nagrani17,
  author       = "A. Nagrani*, J. S. Chung*, A. Zisserman ",
  title        = "VoxCeleb: A Large-scale Speaker Identification Dataset.",
  booktitle    = "INTERSPEECH, 2017",
  year         = "2018",
}