Source-linked AI summary
Opencpop: A High-Quality Open Source Chinese Popular Song Corpus for Singing Voice Synthesis
Yu Wang, Xinsheng Wang, Pengcheng Zhu, Jie Wu, Hanzhao Li, Heyang Xue, Yongmao Zhang, Lei Xie, Mengxiao Bi
TL;DR
Mandarin SVS lacked an accessible, high-quality public corpus, partly because singing requires labor-intensive alignment and musical-score annotation. The paper releases Opencpop, a manually annotated 100-song corpus, and evaluates baseline SVS models; the best model achieves 3.70 MOS, supporting the corpus's reliability.
Problem
Mandarin SVS lacked an accessible public high-quality corpus, while aligning singing with musical scores requires substantial professional manual post-processing.
Method
The paper constructs Opencpop from 100 Mandarin Pop songs with studio recordings, manual phonetic and musical-score annotation, and baseline SVS models for evaluation.
Results
3.70 MOS was achieved by CpopSing trained on Opencpop, indicating the database's reliability.
Takeaways & Limitations
Opencpop provides an open accessible high-quality Mandarin singing corpus with manual annotation for further SVS research.
Takeaways & Limitations
The corpus is performed by a professional young female singer, defining its primary vocal coverage around that singer and selected songs.
Abstract
from arXiv · showhide
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.
1. Introduction
Opencpop addresses the lack of an accessible, high-quality public Mandarin SVS corpus by releasing manually annotated recordings with musical scores. The paper also builds baseline SVS models to assess the corpus and support future comparison.
- SVS synthesizes singing voices from text and musical score information for applications including virtual avatars and artistic creation.
- No accessible public Mandarin SVS corpus with high quality existed, limiting Mandarin SVS development despite prior research and released databases.
- High-quality SVS corpus construction requires musical-score annotations because singing does not align perfectly with the score, creating substantial professional post-processing work.
- Opencpop contains 100 Mandarin Pop songs recorded by a professional female singer in a studio at 44,100 Hz, totaling around 5.2 hours.
- Baseline SVS models are trained on Opencpop to test the database's reliability and provide baseline performance for future research.
2. The creation of Opencpop
The paper presents Opencpop as a publicly available Mandarin singing corpus for SVS, with manually annotated textual information and musical scores. Its creation process is documented through a flowchart and detailed procedural descriptions.
- Opencpop is designed for singing voice synthesis and provides fine manually annotated textual information and musical scores.
- The corpus creation workflow is summarized in Figure 1 and elaborated through descriptions of each construction step.
2.1. Songs and the singer
The corpus uses 100 Mandarin Pop songs selected to provide sufficient phoneme and BPM coverage, performed by a professional young female singer. The singer's vocal range covers all selected songs.
- 100 Mandarin Pop songs with sufficient phoneme and BPM coverage were selected for recording.
- The recordings feature a professional young female singer whose vocal range can perform all selected songs.
2.2. Recording
Recordings were made as isolated vocals in a professional studio, with headphone accompaniment used to support steady and accurate singing. The labeled sample representation includes phoneme examples and rest annotations for silence or aspirate.
- Singing was recorded as a pure vocal track in a low-reverb professional studio without background music.
- Headphone-provided accompaniment was used during recording to ensure the singer's steadiness and accuracy.
- In the labeled note track, “rest” represents either silence (SP) or aspirate (AP).
2.3. Labeling
Opencpop constructs musical-score and phonetic annotations from recorded singing rather than relying solely on the original score. The process combines automatic tools with manual correction and aligns note boundaries with syllable boundaries.
- MIDI annotation: The original score was insufficient for exact alignment, so the MIDI score was created from recorded audio and refined manually.Professional annotators adjusted the automatically generated preliminary score before further tuning during TextGrid annotation.
- TextGrid annotation: TextGrid annotation labels note pitch, note duration, phoneme boundaries, and syllable boundaries across annotation levels.Note durations are fine-tuned to align with newly labeled syllable boundaries.
- TextGrid annotation: Montreal Forced Aligner provides preliminary phoneme–audio alignment that annotators refine in Praat.Syllable boundaries are derived from phoneme boundaries, while pitch boundaries are adjusted to match syllable boundaries.
- TextGrid annotation: Opencpop annotates slurred phoneme repetitions and aspirates in addition to ordinary phoneme and syllable boundaries.For example, a phoneme may repeat within a syllable because of a slur.
2.4. Post processing
After annotation, Opencpop segments recordings into sentence-level utterances for SVS training and further splits utterances longer than 8 seconds. The test set contains 5 songs and 206 utterances, while 95 songs and 3,550 utterances form the training set.
- Segmentation: Audio is first segmented into sentence-level units using transcription and annotated boundaries.Utterances longer than 8 seconds are split further, prioritizing the end boundary of an internal rest.
- Train/test split: 206 utterances from 5 songs constitute the test set, while 3,550 utterances from 95 songs constitute the training set.The selected test songs cover relatively high, low, and medium pitch cases for evaluation.
2.5. Statistics of Opencpop
Opencpop spans varied tempo, utterance duration, pitch, and Mandarin phoneme coverage. Most songs fall within 58–94 BPM, while notes mainly occupy MIDI 56–70 and all Mandarin phonemes are represented.
- BPM: 40 to 130 BPM: Opencpop includes songs across a broad tempo range, with most songs between 58 and 94 BPM.The 58–94 BPM interval is described as common for Mandarin pop music.
- Utterance duration: 1.8s and 8.0s: the shortest and longest segmented utterances define the observed duration limits.Most utterances range from 3 to 7 seconds.
- Pitch: MIDI note 56 to 70: the notes are mainly distributed from G#3 to B4.These correspond to approximately 196 Hz and 494 Hz, respectively.
3. Baseline systems and experiments
The experiments establish baseline SVS systems on Opencpop using a two-stage acoustic-model/vocoder framework and compare Transformer- and Conformer-based variants with objective and subjective metrics. CpopSing achieves 3.70 MOS, while the results also expose difficulty with long-tail pitch distributions.
- Methods: The baseline framework combines Fastspeech2 for spectrogram synthesis with HiFi-GAN for waveform reconstruction.This is a two-stage singing voice synthesis architecture.
- Methods: The SVS Fastspeech2 variant concatenates phoneme and note embeddings and adds note-level duration supervision.Its duration predictors also use note duration, slur indicators, phoneme position, and phoneme count within a note.
- Methods: CpopSing replaces Transformer blocks with Conformer blocks and adds multi-length spectrogram adversarial training alongside L1 loss.The design targets fine-grained local context and over-smoothed acoustic features.
- Results: CpopSing achieves 3.70 MOS on Opencpop, indicating the corpus supports reliable SVS evaluation.The score is near the paper’s stated good-quality range.
- Results: CpopSing and CFSing outperform TFSing on all metrics except duration accuracy, with similar objective performance but higher subjective quality for CpopSing.The evaluation uses F0-RMSE, F0-PCC, duration accuracy, and listener ratings; each audio is rated by 20 listeners.
- Results: Baseline models show limited ability to handle pitches in the training set’s long-tail frequency distribution.The paper identifies pitches beyond the training distribution as a future challenge.
4. Conclusions
The paper introduces Opencpop as an open-access Mandarin singing corpus for SVS with manual phonetic and note-boundary annotation. The authors present it as a resource for further SVS research.
- Opencpop is designed primarily for singing voice synthesis systems.
- All recordings are manually annotated with phoneme boundaries and note boundaries.
- The corpus is presented as an open accessible high-quality Mandarin singing corpus that can support further SVS research.