Opencpop: A High-Quality Open Source Chinese Popular Song Corpus for Singing Voice Synthesis

Abstract

This paper introduces Opencpop, a publicly available highquality Mandarin singing corpus designed for singing voice synthesis (SVS). The corpus consists of 100 popular Mandarin songs performed by a female professional singer. Audio files are recorded with studio quality at a sampling rate of 44,100 Hz and the corresponding lyrics and musical scores are provided. All singing recordings have been phonetically annotated with phoneme boundaries and syllable (note) boundaries. To demonstrate the reliability of the released data and to provide a baseline for future research, we built baseline deep neural network-based SVS models and evaluated them with both objective metrics and subjective mean opinion score (MOS) measure. Experimental results show that the best SVS model trained on our database achieves 3.70 MOS, indicating the reliability of the provided corpus. Opencpop is released to the open-source community WeNet 1 , and the corpus, as well as synthesized demos, can be found on the project homepage 2 .

WeSinger 2: Fully Parallel Singing Voice Synthesis via Multi-Singer Conditional Adversarial Training

“…In addition, for some rare and difficult musical notes, WeSinger 2 also demonstrates a much better performance than WeSinger [10], which can be reflected in the lower F0 RMSE value. Besides, we found that most of the generated singing audios by WeSinger 2 do not have metallic or inconsistent artifacts as found in [30] and do not sound unnatural. Apart from the performance, around 3 times faster than real-time synthesis can be achieved on a single core of a moderate CPU @ 2.60 GHz in Python environment.…”

“…The height of each randomly selected region represents the frequency band and the width represents the number of frames. We fix the values of height to be [20,30,40,50] and the values of width to be [190,160,70,30], which can effectively capture different resolutions in both the time domain and frequency domain. The location of each rectangle region is randomly sampled at every training step to achieve the effect of data augmentation.…”

“…• Opencpop [30], a public female singing dataset 2 including 3550 segments for training and 206 segments for testing.…”

Detecting Voice Cloning Attacks via Timbre Watermarking

“…Thus, we also try our method in such a scenario and validate its effectiveness. Specif- ically, we use the first 30 singing voices (around 5s per voice) from the Opencpop dataset [74], embed watermarks in them, and then use these voices to train so-vits-svc according to the instruction document. After that, we use the trained model to perform voice conversion on a song.…”

Transcription of Mandarin singing based on human-computer interaction

“…The original datasets mainly include the Opencrop dataset [9] released by Yu W et al, the OpenSinger [10] dataset released by Huang R et al, and the PopCs [11] dataset released by Liu J et al The detailed introduction of these datasets is as TABLE 1.…”