Variable Speed Speech Synthesis

We propose a completely trainable speech synthesis solution that is capable of synthesizing variable speed speech from different speakers under various speaking status. Specifically, our proposed system consists of a speaking status synthesis module, a speaker (accent) morphing module and a text-to-speech synthesis module. The novelty and uniqueness of our proposed approach are as follows: First, it decouples speaking status and speaker voice characteristics, and model them separately; second, it comprises two completely trainable modules. This allows the system to synthesize a variety of speech, e.g. fast speech under highly stressful condition.