Quality improvement of HMM-based synthesized speech based on decomposition of naturalness and intelligibility using non-negative matrix factorization

Hidden Markov model based synthesized speech is intelligible but not natural because of over-smoothing of the speech spectra. The purpose of this study is improving naturalness without violating acceptable intelligibility by decomposing the naturalness and intelligibility of synthesized speech using a novel asymmetric bilinear model involving non-negative matrix factorization. Subjective evaluations carried out on English data confirm that the proposed method outperforms original asymmetric bilinear model involving singular value decomposition in factorizing naturalness and intelligibility. Moreover, the performance of the proposed method is comparable with other methods.