Fix audio loading for API: use torchaudio directly

Add dual F0 visualization: extracted vs predicted

Fix CTC decode: blank_index as positional argument

Fix module access: use self.hparams instead of self.modules

first commit