This is the PyTorch implementation of speaker embedding (d-vector) trained with GE2E loss.
The original paper about GE2E loss could be found here: Generalized End-to-End Loss for Speaker Verification
To use the script provided here, you have to organize your raw data in this way:
- all utterances from a speaker should be put under a directory (speaker directory)
- all speaker directories should be put under a directory (root directory)
- speaker directory can have subdirectories and utterances can be placed under subdirectories
You have to specify two things here:
- use
-sor--save_dirto specify the directory for saving processed utterances - use
-cor--config_pathto specify the path to the configuration for Audiotoolkit module
And you can specify the maximum amount of utterances to be extracted and preprocessed for a single speaker, e.g. -m 50.
And a good thing about this script is that you can extract utterances from multiple root directories. For example:
python prepare.py -s data-dir -c toolkit_config.yaml VCTK-Corpus/wav48 LibriSpeech/train-clean-360Only DATA_DIR, MODEL_DIR and CONFIG_PATH have to be specified here.
For example:
python train.py data-dir model-dir dvector_config.yamlNote that the configuration needed here is different from the one for preprocessing.
For more details, check the usage with python train.py -h.
During training, event logs will be put under MODEL_DIR.
To continue the training from a saved checkpoint, just specify the checkpoint path with -c or --checkpoint_path.
Note that you can still specify other optional arguments because they might be different from the ones in the previous training.
The dimension reduction result (using t-SNE) of some utterances from LibriSpeech.
The GE2E-Loss module is first borrowed from cvqluu/GE2E-Loss and then rewritten and optimized for speed by myself.