Pretrained Evaluation Models

This release contains evaluation models pretrained on non-anonymized (original) LibriSpeech-train-clean-360. They consist of two models for ASV (one for xvector encoder, one for ECAPA-TDNN) and one ASR model.

To be in accordance with the provided config files, create a directory exp in the top directory of this repository and place the unzipped folders there. Alternatively, change the model paths in the config files to a different location. Then, we also need access to the ASR models from the asr subfolder. For this, either create a symlink between the subfolder and the exp directory:

ln -srf evaluation/utility/asr/exp exp

or create an exp directory manually in the asr subfolder and provide the symlink directly to the asr model:

mkdir evaluation/utility/asr/exp
ln -srf evaluation/utility/asr/exp/asr_pre exp/asr_pre

All newly created models within the framework should have the symlink added automatically.

In the end, the structure should look like follows:

VoicePAT
    └─ evaluation
        └─ utility
            └─ asr
                └─ exp
                    └─ asr_pre
    └─ exp
        └─ asr_pre
        └─ asv_pre_ecapa
        └─ asv_pre_xvector