Disentanglement of speaker from speech by combining LORD and AdaIN-VC.
For all demonstrated files, using --help shows documentation both at file level and command level (e.g. main.py --help lists commands, main.py train --help details available flags).
For training, any key in the dictionary in config.py can be overridden. Nested keys can be accessed using a /, e.g. --train_encoders/n_epochs 200.
Preprocessing a dataset:
main.py preprocess <data_dir> <save_dest> [<segment>]
Training a model with latent optimization (first stage):
main.py train <data_path> <save_path>
Training the autoencoder (second stage):
main.py train_encoders <data_path> <model_dir>
Converting the speaker of an audio sample to that of another using a trained model:
inference.py convert <model_path> <content_file_path> <speaker_file_path> <output_path>
Multiple conversions can be performed at once by creating a file with lines of the form <content_file_path> <speaker_file_path> <output_path> which we call a pairs file:
inference.py convert_many <model_path> <pairs_file_path>
Calculating neural MOS scores for generated samples:
evaluation.py mean_opinion_score <data_dir>
Evaluate the similarity between the speakers in converted files and their original recordings:
evaluation.py speaker_verification <converted_files_dir> <speakers_dir>
Creating t-SNE plots for class and content:
evaluation.py tsne <data_dir> <model_path> [<segment>] [<n_utterances>]