N-Singer : Non-Autoregressive Korean Singing Voice Synthesis

System for Pronunciation Enhancement

Gyeong-Hoon Lee, Tae-Woo Kim, Hanbin Bae et al

{ghlee3401, ktw0114, bhb0722}@ncsoft.com

Abstract

    We propose N-Singer, a fast and robust non-autoregressive Korean singing
voice synthesis system that can synthesize high-quality and articulate pronounced
Korean singing in parallel. N-Singer consists of a transformer-based mel-generator,
a convolutional network-based postnet, and voicing-aware discriminators: 1) For
accurate pronunciation, N-Singer separately models linguistic and pitch information
without other acoustic features, e.g. mel-spectrogram. 2) To achieve improved mel-
spectrograms, N-Singer uses a combination of transformer-based modules and con-
volutional network-based modules. 3) In the adversarial training, we use voicing-
aware conditional discriminators to capture harmonic features of voiced segments
and noise components of unvoiced segments respectively. The experimental results
prove that N-Singer can synthesize a natural singing voice in parallel with more
accurate pronunciation than the baseline.

Contents

1. Audio Samples (Korean)