Hello,

If we use greedy decoding without sampling, the model cannot generate the speech well.
Sometimes the inference is ended earlier without finishing the full sentence, or generating random noises.
It seems that there is no error in training and generation code, do you have any insight about this behavior?