A pytorch implementation of MBNET: MOS PREDICTION FOR SYNTHESIZED SPEECH WITH MEAN-BIAS NETWORK. Please notice that although this repo reproduces the results in the paper, this is unofficial.
To train a new model, please run train.py, the input arguments are:
- --data_path: The path of the directory containing all .wav files of VCC-2018 and
the train/dev/test split files (the files in
./data). - --save_dir: The path of the directory to save the trained models. Please create the directory before training.
- --total_steps: The total #training step in the training.
- --valid_steps: Do the validation every #(valid_steps) of training update.
- --log_steps: Log the tensorboard every #(log_steps) of training update.
- --update_freq: Gradient accumulation, the default value is 1 (no accumulation).
To test on VCC-2018, please run test.py, the input arguments are:
- --model_path: The path to the saved model.
- --idtable_path: The path to the "judge id-number" mapping table file used during training.
- --step: The time step for tensorboard log, which can be the same as the training steps.
- --split: The valid/test split of data to be used in the testing.
After training on the VCC data, the model can be utilized to inference on other data. The input arguments are --data_path, --model_path, --save_dir, which are similar to the above. Notice that the bias-net is not used since in this code the ground-truth judge ids are assumed to be unavailable.
The pre-trained model can be found in ./pre_trained.