Recently, Sun Yanzi's voice has been used for various covers, and I also tried it myself, covering a small section of the song "Under the Fuji Mountain", for entertainment and learning purposes.
Tools:
I referenced some tutorials from experts and used pre-trained models. The main tools used were:
- URV5: to cut the song to be replaced
- Sun's voice model: can be trained by oneself, but here I directly used a pre-trained model
- so-vits-svc: for training and converting vocals (web UI version from Bilibili user "Feather Quilt")
Some thoughts:
The field of AI still cannot avoid copyright issues. Generally, timbre is not protected by copyright because it is difficult to give a unique identifier to timbre. Just like this AI cover of a song, it can only be said to sound very similar to Sun's voice. As long as it is not acknowledged, it is impossible to define the act of using Sun's timbre. It is even possible to actively adjust the AI's timbre to achieve 90% similarity, making it completely indistinguishable.
Copyright protection applies to works, such as songs and articles, which are content creations. This is something to consider when commercializing AI voices. Either buy the copyright or create your own.
The application scope of AI voices is very broad and suitable for creating audio content. One approach is to use existing well-known voices and blend them with different vocal styles. For example, using Sun's timbre but pronouncing certain words in the style of Jay Chou. In terms of content, AI-generated text can be used with the help of AIGC tools. Finally, it is important to consider the application scenarios and creativity.
I already have some initial ideas and will try them out.