Some notes about Voice Sythesis
Some notes for voice synthesis tutorial videos I looked these days:
For vits:
· Training in Linux has better efficiency and less bug when compared with Windows.
· Emotional Model https://github.com/audeering/w2v2-how-to
· Tell the machine what emotion is, Chinese: https://github.com/innnky/emotional-vits (Use cluster to simplify multi-models work)
· For the emotional model, vits can synthesize a sentence close to the target sentence, but it doesn’t know what the emotion means, it’s just a parameter.
· Environment and requirements, Chinese: https://www.bilibili.com/video/BV1zy4y197Tz/?spm_id_from=333.788.top_right_bar_window_history.content.click&vd_source=39f2b3e527aec166fc87c51a6e6273f8
· Assumption: It seems in the future, if owning enough samples for one voice actor, just pick out an emotional voice sample as a standard tag to define the specific emotion and provide enough sample tags to classify the emotion. Then this voice actor’s unique emotional expression can be controlled. If there are enough similar tags for different actors, can clustering extract characters from the same emotional expression among different characters?
For Diffsinger:
· Very powerful, and the quality can match the origianl sources from Sythersizer V . The most important part is it is an open-source project.
· It can also simplify the process. After tagging the pronunciation, it needs less hand adjustment.
· Export the song database to the speech database (https://www.bilibili.com/video/BV1XD4y177NY/?vd_source=39f2b3e527aec166fc87c51a6e6273f8)
Comments