Harnessing Latent Variation in DNN-Based Speech Synthesis

期間：2017年4月 – 2018年3月 (中途終了)
助成種目：日本学術振興会 科学研究費助成事業 若手研究(B)
課題番号：17K12720
URL：https://kaken.nii.ac.jp/ja/grant/KAKENHI-PROJECT-17K12720

With this grant, I have derived and published theoretical connections between common (heuristic) practical methods for unsupervised learning of controllable speech synthesisers, and latent variables in Bayesian probability, including how common extensions of the practical approach can be given a probabilistic interpretation. Related work (published as well as submitted) explored the optimal supervised methods for annotating the same data, and (separately) considered speech synthesis with multilingual phonetic control. A listening test is currently comparing the aforementioned supervised and unsupervised approaches against variational autoencoders (VAE) and a journal manuscript with the results, and new theoretical connections between VAE and common synthesis heuristics, is in preparation.