A DNN-based emotional speech synthesis by speaker adaptation

Abstract

The paper proposes a deep neural network (DNN)-based emotional speech synthesis method to improve the quality of synthesized emotional speech by speaker adaptation with a multi-speaker and multi-emotion speech corpus. Firstly, a text analyzer is employed to obtain the contextual labels from sentences while the WORLD vocoder is used to extract the acoustic features from corresponding speeches. Then a set of speaker-independent DNN average voice models are trained with the contextual labels and acoustic features of multi-emotion speech corpus. Finally, the speaker adaptation is adopted to train a set of speaker-dependent DNN voice models of target emotion with target emotional training speeches. The target emotional speech is synthesized by the speaker-dependent DNN voice models. Subjective evaluations show that comparing with the traditional hidden Markov model (HMM)-based method, the proposed method can achieve higher opinion scores. Objective tests demonstrate that the spectrum of the emotional speech synthesized by the proposed method is also closer to the original speech than that of the emotional speech synthesized by the HMM-based method. Therefore, the proposed method can improve the emotion express and naturalness of synthesized emotional speech.