Description

Before Start This version is not reserved for covers with high intonation (which can lead to over-saturation and thus distorted vocals). The dataset is only trained with gawr gura's live content when she sings "softly" with a relatively weak voice. - I'll probably do a version 2 soon, this time with a more "varied" dataset, bearing in mind that currently there's only 2/4 of the original dataset in this version. - After listening to the previews over and over again (or forgetting to stop) for almost 2 hours non-stop, I can no longer decide whether the model is really that good. You tell me Last Update : <t:1703096452:R> - Model URL Version RVC V2.0 - Pitch Extraction Algorithm RMVPE - Epochs - Steps 400 - 12.8 K - Dataset ~ 00:13:50 - Recommended Usage Cover - Search Feature Ratio 0.75 - Pitch Logic Pitch ( +6~+12 = Man / 0 = Women ) - You can adjust if you found an better result - As mentioned above, the model is not designed for singing to music where the singer is very, very high pitched. It should therefore be made to sing on music with a singer who sings softly, or at least who doesn't go very high. You could lower the pitch if this happens, but you might distort the overall result with a voice that's too low. Previews Preview_Cover_1.wav *Contains External Effects* : Yes - *Pitch* : +0 - *Feature Ratio* : 0.75 - Preview_Cover_2.wav *Contains External Effects* : Yes - *Pitch* : +6 - *Feature Ratio* : 0.75