some questions about model inference

[2901733926]

Is the converted audio in the sample obtained using inference-from-mel or inference-from-wav? The paper mentions using hifigan as the vocoder and that hifigan was also trained on the training dataset, but it doesn't seem to be represented in the code.