VoxMovies

This repository contains evaluation scripts and pretrained models for VoxMovies dataset.

project_page arxiv_paper

Dependencies

pip install -r requirements.txt

Dataset and test pairs

• VoxMovies dataset and test pairs can be downloaded from here.
• For e2, e3 and e4 test pairs, you also need both VoxCeleb1 dev and test set. Please download them from here.

Evaluation

• You need to put VoxCeleb1 and VoxMovies into args.test_path. We've used symlink for this.
• Please note that PATH_TO_VOXCELEB1 directory needs to contain both VoxCeleb1 dev & test wavfiles.

mkdir data
cd data
ln -s PATH_TO_VOXCELEB1 voxceleb1
ln -s PATH_TO_VOXMOVIES_TEST_SET voxmovies_test

Then, run the script below.

python eval.py --initial_model PATH_TO_PRETRAINED_MODEL --test_list PATH_TO_TEST_PAIRS --test_path data/

Pretrained models

• Both pretrained model and finetuned model are provided below.
• Note that baseline_v2_ap.model has already been publicly available. Please refer to here for more details about model architecture and training procedures.

Evaluation results (EER %) with VoxMovies test sets

Test pairs	e1	e2	e3	e4	e5
baseline_v2_ap.model	6.09	7.40	7.50	9.23	10.47
finetuned.model	5.76	7.10	8.36	7.37	9.55

Codebase

This evaluation code is largely based on clova voxceleb trainer. Please refer to this repo if you want to train/finetune the model.

Citation

If you make use of this code, kindly reference:

@InProceedings{Brown21b,
    title={Playing a Part: Speaker Verification at the Movies},
    author={Andrew Brown and Jaesung Huh and Arsha Nagrani and Joon Son Chung and Andrew Zisserman},
    year={2021},
    booktitle={International Conference on Acoustics, Speech, and Signal Processing (ICASSP)}
}