hcy71o / snac Goto Github PK

I have reviewed the paper, which asserts that there is no audio or information from unknown speakers during the training process. However, in your code, your validation set includes information from unknown speakers. Doesn't this imply that the model had already assimilated information from unknown speakers during training, and the reference speech used for inference corresponds to the speech encountered during training?

For speaker p261，the model already meet this speaker？

please help me！

请问可以小样本克隆音乐吗？

In my experiments, the wav generated is not similar as the ref audio

안녕하세요. SNAC 논문 잘 봤습니다!
현재 저는 proposed + pre-trained + flow 형태로 만들고 있는데요.
snac관련 부분은 해당 코드를 그대로 가져와서 구현했습니다(loss도 변경완료했습니다)
그런데 kl loss가 계속 nan으로 발산하는데 혹시 추가적으로 이런현상을 제어하기위해 코드변경하신부분이 있으신가요?
아니면 이런 현상을 겪어 보신적이 있으시면 해결방법 알려주시면 감사하겠습니다.
감사합니다.

Hello, are there any intensions to implement "Proposed+REF+FLOW" along with the already implemented "Baseline + REF + FLOW"

Thanks in advance.