Support for the additional voices
Hi, this was great work.
I was wondering how to support additional voices besides the predefined ones. Do we need to fine-tune the model for each new voice (for the English language), or can we perform inference directly using a custom voice?
I tried to create a custom voice for inference by simply providing the model with an audio sample array of shape (1, 256), which the model accepts. I loaded the audio, resampled it to 24kHz, clipped the first 256 elements, and passed it to the model. However, the output was gibberish.
Could you explain how to correctly obtain a style voice sample and pass it to the model? Do we need to use a specific model to encode the audio into the required dimensions, or are there alternative methods for voice styling in this architecture?