Trump voice
aomv opened this issue · 1 comments
Hi,
I’m trying to understand how did you train Donal Trump, Obama, Mark Zuckerberg and Sheryl Sendberg voices.
Could you provide an example of how to train a new voice like you did?
On this article the author mentions "learning new voice embeddings after the model has been trained. To do this they freeze all the other parameters except the new voice embedding (a 256-vector) and train just those weights with new audio data from the new voice. These voice samples comprise “10s of minutes” for each new voice and come from Youtube videos that were transcribed by an automatic speech recognition system (presumably the one built into Youtube). The authors point out that these audio samples are much less uniform and more noisy than the original corpus, and include things like clapping and occasional other speakers” (http://kbullaughey.github.io/lstm-play/2017/10/27/voice-loop-summary.html)
How could I achieve that?
Thank you!