MULT-SPEAKER HELP
Ashwin-Ramesh2607 opened this issue · comments
Hi,
I want to know some details about the multi-speaker training. In specific, what changes are present in the architecture/data preparation for single vs multi-speaker dataset. Are you merging multiple speakers in one audio clip or not. Please explain the steps for training a multi-speaker custom dataset.
There are no architectural differences involved in training a multispeaker model instead of a singlespeaker one. The difference is purely in the dataset. For the singlespeaker task, all of the patches of audio that we train on come from the same speaker. For the multispeaker task, these patches come from many speakers.
To train on a custom dataset, you'll need to create the h5 files that run.py expects. You can use prep_vtck.py as a guide.