HMM-based polyglot speech synthesis by speaker and language adaptive training

This paper describes a technique for speaker and language adaptive training (SLAT) for HMM-based polyglot speech synthesis and its evaluations on a multi-lingual speech corpus. The SLAT technique allows multi-speaker/multi-language adaptive training and synthesis to be performed. Experimental results show that the SLAT technique achieves better naturalness than both speaker-adaptively trained language-dependent (LD-SAT) and language-independent (LI-SAT) models. In cross-lingual adaptation speaker similarity tests SLAT and LI-SAT outperform LD-SAT but there are still significant differences between polyglot adaptation and intra-language adaptation.