SV-eval is a speaker verification evaluation toolkit designed to process and evaluate speaker verification models using various datasets. The toolkit supports multiple models and provides functionalities for data preprocessing, model evaluation, and result analysis.
python src/main.py --model_config=<model_name>Each model configuration is defined in the config directory. The configuration file contains the model hyperparameters, dataset paths, and other model-specific parameters.
Place the datasets in the data directory. The toolkit supports the following datasets:
- VoxCeleb1
- VoxCeleb2 (customized structure)
The toolkit supports the following models:
- ECAPA-TDNN
- CAM++
- ECAPA2
- ReDimNet
The main config is located in root directory.
dataset_name: 'vox1_test_wav'
dataset_path: '../data/'
max_len: 54_000 # 4 * 16000
batch_size: 32
embeddings_output_path: '../embeds/'
results_output_path: '../results/'
dataset_type: 'voxceleb1'The model config is located in ../data/configs directory.
model_name: 'campplus'
fbank_processing: True
device: 'mps'
threshold:If threshold is not provided, the model will calculate threshold to minimize EER.
If threshold is provided, the model will use this threshold to calculate EER.
The results are saved in the results directory. The results include the EER, accuracy, and other evaluation metrics.
