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

TL;DR

Opencpop introduces a high-quality open Mandarin SVS corpus comprising 100 professionally sung songs with studio-grade 44.1 kHz audio and comprehensive MIDI/TextGrid annotations, enabling precise note and phoneme alignment for SVS research. The authors build baseline SVS systems, including a Conformer-based model with adversarial mel training (CpopSing), and demonstrate superior subjective MOS (3.70) relative to Transformer-based baselines, while providing thorough objective metrics. The dataset covers diverse BPMs, pitches, and phonemes, and is split into a 5-song test set to ensure broad evaluation; results validate Opencpop as a reliable benchmark for Mandarin SVS. The work aims to catalyze Mandarin SVS development by offering a publicly accessible resource and baseline references for future research and benchmarking.

Abstract

This paper introduces Opencpop, a publicly available high-quality 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, and the corpus, as well as synthesized demos, can be found on the project homepage.

Figures (7)

• Figure 1: Flowchart for creating the Opencpop database. For clear presentation, here, quality assurance is only illustrated after the TextGrid annotation. In practice, the human double-check exists in every process.
• Figure 2: Illustration of a labeled sample. The phoneme "d" of syllable "dian" and the phoneme "w" of syllable "wang" are omitted due to the limited space. The "rest" in the note track corresponds to the silence (SP) or aspirate (AP).
• Figure 3: The violin plots of pitch numbers in different subsets. Pitch is presented as MIDI note number, where A4=69=440 Hz.
• Figure 4: The statistical distribution of BPM.
• Figure 5: The statistical distribution of utterance duration. The shortest and longest duration is 1.8s and 8.0s, respectively.