CWT spectral loss for training a DNN-based speech waveform model