[国内学会] Stealing your vocal identity from the internet: cloning Obama’s voice from found data using GAN and Wavenet

情報処理学会 第118回音楽情報科学・第120回音声言語情報処理合同研究発表会

With the rise of machine-learning speech processing techniques, the risk of having your vocal identity stolen from online audio clips has risen considerably. In this research we study how GAN-based technologies can allow us to significantly improve the vocal quality of found audios that suffer from noise or reverberation, and the consequences that processing has from the speech generation point of view when using state-of-the-art Wavenet-based voice generation techniques.