pretty_name: CrawlSinger-OS
license: mit
language:
- zh
task_categories:
- text-to-speech
- text-to-audio
- automatic-speech-recognition
size_categories:
- 100K<n<1M
tags:
- audio
- music
- singing
- singing-voice-synthesis
- music-score
- arxiv:2607.27768
CrawlSinger-OS
CrawlSinger-OS is a large-scale, open-source singing corpus constructed for score-native singing voice synthesis. It contains more than 2,300 hours of processed singing data from multiple public song and singing collections, with a unified annotation scheme for lyrics, MIDI pitches, symbolic note values, lyric-to-note alignment, and global tempo.
Why CrawlSinger-OS
Modern singing synthesizers benefit from large and diverse training corpora, but conventional SVS datasets are usually smaller than 100 hours. CrawlSinger-OS processes both synthetic and real-world music with the same pipeline and score representation, providing the scale and fine-grained lyric--note alignment needed to train VocalRender's autoregressive diffusion architecture.
The core CrawlSinger-OS release contains:
| Source | Processed hours | Segments | Type | Original annotation |
|---|---|---|---|---|
| Muse (Chinese subset) | 2,013 | 601k | Synthetic songs | Time-aligned lyrics |
| MuChin | 158 | 84k | Real songs | Time-aligned lyrics |
| SongFormDB (Ext) | 93 | 48k | Real songs | None |
| OpenSinger | 53 | 43k | Real singing | Lyrics |
| Total | 2,317 | 776k | Synthetic + real | Unified below |
Synthetic and real subsets have different musical and semantic distributions. The paper therefore trains on them in two stages instead of treating them as interchangeable data.
Data construction
The SingCrawl processing pipeline has four stages:
- Vocal extraction. Cascaded mel-RoFormer models remove accompaniment and reverberation from song audio.
- Slicing and lyric transcription. Audio is divided into clips shorter than 30 seconds. Depending on the source annotations, the pipeline uses a direct ASR path, a candidate-lyrics context-biasing path, or a refined timestamp path.
- Forced alignment. A retrained SOFA aligner with G2PW produces fine-grained lyric timing. On held-out GTSinger and M4Singer subsets, the aligner reaches a mean IoU of 0.84 and VlabelerEditRatio (50 ms) of 0.092.
- Pitch transcription. ROSVOT transcribes note pitches; realized pitch durations are quantized into symbolic note values and a global BPM reference.
Repository contents
The repository uses two storage layouts. folder_based sources keep per-song
metadata inside their tar archives; json_file sources provide a central
annotations.json plus audio archives.
| Path | Role | Layout | Archive size |
|---|---|---|---|
muse/ |
CrawlSinger-OS core | folder-based, 3 shards | 51.40 GB |
muchin/ |
CrawlSinger-OS core | folder-based | 4.20 GB |
songformdb/ |
CrawlSinger-OS core | folder-based | 2.34 GB |
opensinger/ |
CrawlSinger-OS core | JSON + audio | 16.74 GB |
m4singer/ |
Additional public training corpus | JSON + audio | 11.33 GB |
gtsinger/ |
Additional public training corpus | JSON + audio | 19.36 GB |
opencpop/ |
Evaluation corpus | JSON + audio | 5.03 GB |
manifest.json |
Layout, byte size, and SHA-256 for every shard | JSON | -- |
The complete repository is approximately 110.4 GB. Select only the subsets needed for your experiment.
Annotation format
JSON-based sources contain a list of entries such as:
{
"item_name": "2002000039",
"word": ["你", "是", "我", "SP"],
"pitch": [50, 62, 60, 0],
"note": ["<NOTE_16>", "<NOTE_4>", "<NOTE_DOT_8>", "<NOTE_DOT_16>"],
"pitch2word": [0, 1, 2, 3],
"bpm": 81,
"wav_fn": "segments/wavs/2002000039.wav",
"word_dur": [0.17847, 0.84090, 0.51098, 0.25584],
"pitch_dur": [0.17847, 0.84090, 0.51098, 0.25584]
}
| Field | Description |
|---|---|
word |
Lyric syllables; AP/SP denote non-lyric regions. |
pitch |
MIDI pitch per note segment; 0 denotes a rest. |
note |
Symbolic note-value token per note segment. |
pitch2word |
Maps each note segment to its lyric index and supports melisma. |
bpm |
Global tempo reference. |
wav_fn |
Audio path relative to the extracted source directory. |
word_dur, pitch_dur |
Optional realized durations for visualization and evaluation; they are not VocalRender conditioning fields. |
Download
Install the Hugging Face CLI, then download only the desired source:
# Inspect the release manifest first
hf download pymaster/CrawlSinger-OS manifest.json \
--repo-type dataset \
--local-dir data/CrawlSinger-OS
# Example: download the processed OpenSinger subset
hf download pymaster/CrawlSinger-OS \
--repo-type dataset \
--include "opensinger/*" \
--local-dir data/CrawlSinger-OS
Extract downloaded audio shards with tar -xf. Use the SHA-256 values in
manifest.json to verify large downloads.
Relationship to VocalRender and VocalRender-Pro
The two released models share the same score-native architecture: an interleaved lyric--note representation, continuous AudioVAE latents, and an autoregressive diffusion model that predicts acoustic patches and termination without an explicit duration predictor.
| Model | Main training data | Training strategy | Main observed trade-off |
|---|---|---|---|
| VocalRender | CrawlSinger-OS (>2,300 h), with additional public singing corpora | 40k-step synthetic pretraining, then 20k-step real-data finetuning | Stronger subjective score following (MS-MOS 2.96). |
| VocalRender-Pro | CrawlSinger (>5,600 h of in-house real singing) | 160k steps with a global batch size of 32,768 continuous tokens | Better intelligibility, speaker similarity, naturalness, and OOD robustness; MS-MOS 2.71. |
On Opencpop, VocalRender-Pro reduces WER from 4.44 to 3.88 and improves speaker similarity from 0.922 to 0.929. On CrawlSinger-Eval, WER changes from 4.52 to 4.45 and similarity from 0.919 to 0.926. The paper attributes these gains to the larger amount of real singing and broader singer coverage. Conversely, VocalRender's higher score-following rating may result from more reliable and precise score annotations in its finetuning subset.
Only CrawlSinger-OS and the accompanying public corpora are distributed in this dataset repository. The in-house CrawlSinger training data used by VocalRender-Pro is not included.
Limitations
- Most automatically produced scores are audio-centric transcriptions: they describe realized ornaments and note splits rather than a composer's concise intent-centric score.
- Synthetic and real subsets have noticeably different musical, semantic, and pitch distributions.
- Automatically transcribed scores may contain chromatic fluctuations, uncommon note values, or overly fragmented note sequences.
- The release primarily targets Mandarin singing and can contain errors from separation, ASR, forced alignment, and pitch transcription.
Citation
@article{chen2026vocalrender,
title = {VocalRender: Score-Native Singing Voice Synthesis for Real-World Composition},
author = {Chen, Yukun and Wang, Tianrui and Mu, Zhaoxi and Yang, Xinyu and Chng, EngSiong},
journal = {arXiv preprint arXiv:2607.27768},
year = {2026},
url = {https://arxiv.org/abs/2607.27768}
}
Please also cite the original source datasets used by the subset(s) in your work.