HiFi-VC
Reproduction of a state-of-the-art voice conversion system.
Project GitHub. Year: 2023
The project aimed on the reproduction of HiFi-VC voice conversion system. Due to absence of open-source code for the model, the missing details were recovered using backward engineering of the TorchScript authors’ checkpoint.
Authors of HiFi-VC released the scripted version of their model that is suitable only for system inference. In the DLA Course Seminar, this checkpoint was analyzed using its .graph, .code, and .parameters variables to restore model’s workflow and hyperparameters.
After obtaining the missing details, the model reproduction was straightforward.