1. download the dataset
2. create the folder named "train_audio_<eng/ger/per>/<IEMOCAP/EMODB/SHEMO>"
3. run the Data_Processing.ipynb to reorganize the datasets to the following structure:

   - train_audio_<eng/ger/per> 
      - <IEMOCAP/EMODB/SHEMO>
          - emotion A
              - wav       
                  - audio A_1        
                  - audio A_2            
          - emotion B
              - wav       
                  - audio B_1         
                  - audio B_2
          ...
              

4. run the command

python encoder_preprocess.py ./train_audio_<eng/ger/per>

5. The output folder is named SV2TTS, which can also be found in the drive:
   • IEMOCAP SV2TTS: https://drive.google.com/file/d/1AecA1KGuIW8t0czRmzvxI0gC7zsZ9joe/view?usp=sharing
   • EMODB SV2TTS: https://drive.google.com/file/d/1NEQx8T5jgRKbKGYZRgKKT9eygS1MrH54/view?usp=sharing
   • SHEMO SV2TTS: https://drive.google.com/file/d/1MNt_JEFmpOqG9xWgjUrrDF7q9095b5pE/view?usp=sharing

Note: In order to lost the data during processing, I changed the partials_n_frames in encoder/params_data.py. Currently don't know what would be affected.

1. will need to download the visdom package to visualize the training process. And start the visdom server before running the command.
2. change the parameter speakers_per_batch equals to the number of emotion classes in encoder/params_model.py. utterances_per_speaker can be tuned.
3. run the command:

python encoder_train.py <my_run> <datasets_root>/SV2TTS/encoder

<my_run> is like a session name, can be defined as any name. And the training process can be seen in the visdom port.

visdom can be disabled using --no_visdom

There is a pretrained model in the saved_models/default folder if don't want to train from scratch.

run the command:

python generate_embed.py

Can use the following options:

• -s: the path to the source audio (e.g. ./train_audio_eng/IEMOCAP)
• -d: the path to the destination embedding.