BMS voices
-
Hi all,
About a year ago, I started experimenting with training voice models based on the existing BMS voices. @Tomcattwo did similar work and shared it with us here.
Unfortunately, I lost access to the GPU cluster I was using at the time, so I put the project on hold. However, I finally got some new GPU time since a few weeks, so I picked up where I left off. As a first experiment, I trained the first pilot voice from BMS on StyleTTS2. This was literally the first attempt, but the results were encouraging enough that I wanted to share them with you.
Hello world
Ghostrider
MaddogsThere’s still a lot to improve. I think the model is overfitting because of the limited training data. Training a multispeaker model might help with that. Also, there are a bunch a variables I can still play with for training and inference, so the results can definitely be finetuned further.
@Tomcattwo, let me know if you want to work on this together.
@Mav-jp, need any new frags? We can also make full sentences now, if that’s easier.
@Boxer, we’ve had email contact a year ago about this, sorry it took so long.
-
nice one, the two last links seem not to work.
So it’s also a thing for training missions. Quite interesting.
-
@Scorpion82 said in BMS voices:
nice one, the two last links seem not to work.
Oh sorry, it seems to be a problem with Onedrive. I’ve updated the links and sometimes refreshing the page after clicking on the link also works. If it doesn’t, I’ll move the files somewhere else for hosting.
So it’s also a thing for training missions. Quite interesting.
Yes, could be used for anything.