Voice source modelling using deep neural networks for statistical parametric speech synthesis

This paper presents a voice source modelling method employing a deep neural network (DNN) to map from acoustic features to the time-domain glottal flow waveform. First, acoustic features and the glottal flow signal are estimated from each frame of the speech database. Pitch-synchronous glottal flow time-domain waveforms are extracted, interpolated to a constant duration, and stored in a codebook. Then, a DNN is trained to map from acoustic features to these duration-normalised glottal waveforms. At synthesis time, acoustic features are generated froma statistical parametricmodel, and from these, the trained DNN predicts the glottal flow waveform. Illustrations are provided to demonstrate that the proposed method successfully synthesises the glottal flow waveform and enables easy modification of the waveform by adjusting the input values to the DNN. In a subjective listening test, the proposed method was rated as equal to a high-quality method employing a stored glottal flow waveform.