RawNet3 is hosted via two repositories. Inference of any utterance with 16k 16bit mono format and Vox1-O benchmark is supported in this repository.
Training recipe, on the other hand, will be supported in https://github.com/clovaai/voxceleb_trainer.
Model weight parameters are served by huggingface at https://huggingface.co/jungjee/RawNet3, which is used as a submodule here
To download the model, run:
git submodule update --init --recursive
Run: python inference.py --inference_utterance --input {YOUR_INPUT_FILE}
Optionally, --out_dir can be set to direct where to save the extracted speaker embedding. (default: ./out.npy)
Run: python inference.py --vox1_o_benchmark --DB_dir
Note that DB_dir should direct the directory of VoxCeleb1 dataset.
For example, if DB_dir="/home/abc/db/VoxCeleb1",
VoxCeleb1 folder is expected to have 1,251 folders inside which corresponds to 1,251 speakers of the VoxCeleb1 dataset.
If you successfully run the benchmark, you will get:
Vox1-O benchmark Finished. EER: 0.8932, minDCF:0.06690.