HiPhi-GAN

Abstract

Most modern neural vocoders generate a waveform from a mel-spectrogram, one of the acoustic features. Mel-spectrogram is information about the only magnitude, not phase constant. In other words, all mel-spectrograms obtained from x_ϕ ̅ are the same for all ϕ, when x_ϕ is phase transform by ϕ from the waveform x. Conversely, a neural vocoder that generates a waveform with only mel-spectrogram as input is confused in training because x_ϕ can be ground-truth for all ϕ. In this paper, we propose a universal vocoder consisting of a stage to avoid confusion by fixing ϕ to ϕ ̅ to guarantee the uniqueness (x_ϕ ̅ ) of ground-truth, and a stage to generate a full-band waveform according to the fixed ϕ ̅. Each stage is named phase synchronizer and waveform upsampler. The proposed neural vocoder HiPhi-GAN solves all the existing problems: slow inference speed, lousy audio quality at mid-high-band, and frequent phasing errors.