VOCALOID VOCALOID6
- Thread starter Rylitah
- Start date
Guess we're getting a whisper parameter for AI vbs soon
I don't think they take advantage any specialty hardware acceleration. (Too bad, because Apple ships ML accelerator hardware on their systems.) They're very fast to render though. Usually I edit a note, and the little render progress zooms by in about a half second to a second and it's ready to play.
Typically you need beefy hardware to train models, but just running inference on them is fairly resource-light.
I'm actually really curious about how the V6 engine works under the hood, and I wish they'd publish a paper or something like they did for the original Vocaloid. The existence of vocalo-changer and the relative file size of the editor seem like a hint...I wonder if there's a set of internal "carrier" samples that the editor uses, like a traditional sample based bank, and then it alters the timbre with the voice bank's ML model. Or if the editor phoneme/pitch signals just provoke the model to emit whatever it it's been trained as an "ah" sound or whatever.
Typically you need beefy hardware to train models, but just running inference on them is fairly resource-light.
I'm actually really curious about how the V6 engine works under the hood, and I wish they'd publish a paper or something like they did for the original Vocaloid. The existence of vocalo-changer and the relative file size of the editor seem like a hint...I wonder if there's a set of internal "carrier" samples that the editor uses, like a traditional sample based bank, and then it alters the timbre with the voice bank's ML model. Or if the editor phoneme/pitch signals just provoke the model to emit whatever it it's been trained as an "ah" sound or whatever.