Autoformer

Speaker Embedding

Train your own Speaker Embedding here or use my pre-trained model with LstmDV

The pre-trained model is trained by 913 speaker with 53 utterances , Download the dataset from openSLR train-clean-360.tar.gz and ignore the speaker wich utterances number is lower than 50, the model performance is test with 40 speaker from VCTK dataset.

Model	LstmDV	MetaDV
EER	3%	2%

Data Prepare

• Put your Speaker Embedding model in ./model/static/model.pt

• Run make_spec.ipynb and make_metadata.ipynb with the data as following format.

   - model
     - static
       - model.pt
   - make_data
      - factory
      - wavs
          - 225 (include many audio data)
          - 226
          - ...
          - ...
      - make_metadata.ipynb
      - make_spec.ipynb
  

• After that you will get a ./spmel (default name) folder and a train.pkl, copy ./spmel to root dir.

Training

Available Model

• AutoVC (Original Implement)
• MetaPool
• MetaConv

For Original Training

python train.py --model_name=AutoVC --data_dir=spmel --save_model_name=model_name

For Training with Discriminator

python train_with_discriminator.py --model_name=AutoVC --data_dir=spmel --save_model_name=model_name

For Training with Cycle Gan

## Still Working