Many-to-Many Voice Conversion based Feature Disentanglement using Variational Autoencoder

Voice conversion is a challenging task which transforms the voice\ncharacteristics of a source speaker to a target speaker without changing\nlinguistic content. Recently, there have been many works on many-to-many Voice\nConversion (VC) based on Variational Autoencoder (VAEs) achieving good results,\nhowever, these methods lack the ability to disentangle speaker identity and\nlinguistic content to achieve good performance on unseen speaker scenarios. In\nthis paper, we propose a new method based on feature disentanglement to tackle\nmany to many voice conversion. The method has the capability to disentangle\nspeaker identity and linguistic content from utterances, it can convert from\nmany source speakers to many target speakers with a single autoencoder network.\nMoreover, it naturally deals with the unseen target speaker scenarios. We\nperform both objective and subjective evaluations to show the competitive\nperformance of our proposed method compared with other state-of-the-art models\nin terms of naturalness and target speaker similarity.\n