VGGVox Documentation

The VGGVox verification models map voice spectrograms to a compact Euclidean space. In this space, distances between embeddings directly correspond to a measure of speaker similarity. These embeddings can be utilized for several downstream tasks:

• Speaker verification
• Clustering
• Diarisation

To use the VGGVox models, you must first have Matlab and Matconvnet installed and compiled. The recommended installation method is using the vl_contrib package manager.

1. Install and compile Matconvnet.
2. Use vl_contrib to install and setup the VGGVox package:

vl_contrib install VGGVox
vl_contrib setup VGGVox

After installation, you can evaluate the models directly on .wav audio files using the provided demo scripts. There are three available demos:

• demo_vggvox_identif: For identification models trained on VoxCeleb1.
• demo_vggvox_verif: For verification models trained on VoxCeleb1.
• demo_vggvox_verif_voxceleb2: For verification models trained on VoxCeleb2 (ResNet based).

If you prefer to download the Matconvnet models directly instead of using the package manager, use the following links:

• Identification (VoxCeleb1): vggvox_ident_net.mat
• Verification (VoxCeleb1): vggvox_ver_net.mat
• Verification (VoxCeleb2 - ResNet based): ver_net.mat