NSF-Pupu-Vocoder: A High-Fidelity and Light-Weight
Pitch-Controllable Neural Vocoder

Abstract

Tone shift, the technique of generating singing voices at a target pitch while using desired vocal techniques from other registers, is essential in singing voice synthesis (SVS). To achieve this, SVS systems often rely on pitch-controllable neural vocoders, whose performance remains limited by three factors: complex input feature requirements that hinder SVS integration, sub-optimal pitch controllability, and limited synthesis fidelity. This paper proposes NSF-Pupu-Vocoder to address these issues. In contrast to previous models that rely on complex preprocessing and training schemes, our model uses highly compatible mel-spectrograms and pitch contours as inputs, and achieves feature decoupling simply by injecting a pitch excitation into the exact layer whose temporal resolution can accommodate the pitch. Experiments show that our model achieves competitive pitch controllability and superior synthesis fidelity, while preserving the original singer's timbre during pitch manipulation.