Voice conversion method, model training method, device, medium, and program product

A model training method, used for an electronic device, comprising:

• a training speech obtaining step comprising obtaining multiple training speeches having a second voiceprint;
• a training representation vector determining step comprising determining a training content representation vector and a training voiceprint representation vector corresponding to each training speech, wherein the training content representation vector is obtained based on the speech waveform of the training speech;
• a training speech conversion step comprising inputting the training content representation vector and the training voiceprint representation vector corresponding to each training speech into a preset speech generation network, to obtain a corresponding training converted speech;
• a first calculation step comprising calculating a value of a first loss function based on the training converted speech;
• a second calculation step comprising inputting each training speech and the training converted speech corresponding to each training speech into a preset discriminator network to perform discrimination, and calculating a value of a second loss function;
• a parameter updating step comprising updating network parameters of the speech generation network based on the value of the first loss function, and updating network parameters of the discriminator network based on the value of the second loss function;
• an iterative training step comprising performing iterative training based on the speech generation network and the discriminator network after the parameters are updated, until the value of the first loss function and the value of the second loss function both meet a model convergence condition;
• and
• a generator determining step comprising using the trained speech generation network as a speech generator.