Noisy-to-Noisy Voice Conversion Under Variations of Noisy Condition

Voiceconversion (VC) refers to the transformation of speaker identity a speech target one without altering linguistic content. As recent VC techniques have made significant progress, implementing them in real-world scenarios is also considered, where data some inevitable interferences, most common which are background sounds. On other hand, sounds informative and need be retained applications, such as movies/videos. To address these issues, we proposed noisy-to-noisy (N2N) framework that does not rely on clean models noisy directly by using noise conditions. Previous experimental results proven its effectiveness. In this article, further improve performance introducing pre-trained noise-conditioned model. Moreover, explore impacts conditions, more realistic situations evaluated training set possesses speaker-dependent The demonstrate effectiveness pre-training strategy degradation under strict We then augmentation method overcome limitation. Further experiments showed method.