The evaluation scripts are based on external tools (the VoicePrivacy Challenge 2022, SpeechBrain, ESPnet) but have been modified to fit into this framework and include the proposed improvements.
All scripts regarding the ASV training in privacy/asv/asv_train, including parts of privacy/asv/asv.py, are based on the VoxCeleb recipe by SpeechBrain and adapted for LibriSpeech and this framework.
The additional metrics in privacy/asv/metrics are based on scripts in the framework for the VoicePrivacy Challenge 2022. They were included from other sources, i.e.,
- ZEBRA: https://gitlab.eurecom.fr/nautsch/zebra
- Cllr and linkability: https://gitlab.inria.fr/magnet/anonymization_metrics
The ASR scripts in utility/asr are all based on the LibriSpeech ASR1 recipe by ESPnet. Some scripts were copied from that source verbatim because they are not part of the ESPnet Python library.
The scripts for computing Gain of Voice Distinctiveness (GVD) in utility/voice_distinctiveness are based on the computation of similarity matrices in the VoicePrivacy Challenge 2022.
This documentation is still under construction and will be extended soon.