Research map / Image generation / Autoregressive and latent generation
GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GTSinger, a 80.59-hour, 9-language, 20-singer singing corpus with six annotated singing techniques, realistic music scores, alignments, style labels and paired speech, plus four benchmarks.
Problem
Existing singing datasets are low quality, limited in languages and singers, lack multi-technique controlled annotations, realistic music scores, and task coverage, hindering controllable and personalized singing tasks.
Approach
A large, free-to-use multilingual singing corpus with controlled technique/control group recordings, phoneme-level technique annotations, realistic (regular-duration) music scores, manual alignments, global style labels and paired speech, validated on four benchmark tasks.
Key results
- StyleSinger achieved the best technique-controllable SVS results (parallel FFE 0.25, MCD 3.27, MOS-Q 4.01, MOS-C 4.15), but still below GT (vocoder) MOS-C 4.42.
- Technique recognition reached overall F1 of 0.96 for falsetto and 0.99 for breathy, but only 0.70 for vibrato and glissando.
- StyleSinger outperformed AlignSTS on speech-to-singing (MOS-Q 3.83 vs 3.68; MOS-S 3.88 vs 3.73).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RMSSinger instead of DiffSinger for technique-controllable SVS | Technique-controllable SVS · Parallel | MOS-C: 3.83 $\pm$ 0.07 -> 3.95 $\pm$ 0.05 | Helped |
| use StyleSinger instead of RMSSinger for technique-controllable SVS | Technique-controllable SVS · Parallel | MOS-C: 3.95 $\pm$ 0.05 -> 4.15 $\pm$ 0.06 | Helped |
| use StyleSinger instead of RMSSinger for style transfer | Style transfer · Parallel | MOS-Q: 3.70 $\pm$ 0.04 -> 3.95 $\pm$ 0.06 | Helped |
| use StyleSinger instead of RMSSinger for style transfer | Style transfer · Parallel | Cos: 0.88 -> 0.93 | Helped |
| use StyleSinger instead of AlignSTS | Speech-to-singing · Speech-to-singing | MOS-Q: 3.68 $\pm$ 0.12 -> 3.83 $\pm$ 0.09 | Helped |
| train on one language group and test on the other rather than use the overall split | Technique recognition model · Cross-lingual technique recognition | F1: 0.78 -> 0.72 | Hurt |
Surprising findings
- Cross-lingual technique recognition and style transfer show a notable drop in performance relative to in-domain settings, indicating limited generalization.
Limitations
- Does not cover all widely spoken languages (e.g., Arabic) or techniques such as vocal fry.
- Manual phoneme-duration segmentation and subtle technique identification remain challenging for human annotators.
- Dataset is singing without accompaniment and needs vocal-to-accompaniment models for broader music use.
Experimental setup
Datasets: GTSinger
Benchmarks: GTSinger technique-controllable SVS test set GTSinger technique recognition test set GTSinger style transfer test set GTSinger speech-to-singing test set
Metrics: MOS-Q MOS-C MOS-S FFE MCD Cos F1 Accuracy Precision Recall
Baselines: DiffSinger RMSSinger StyleSinger AlignSTS
Builds on: DiffSinger StyleSinger ROSVOT RMVPE Montreal Forced Aligner
Closest papers by meaning
- TechSinger: Technique Controllable Multilingual Singing Voice Synthesis via Flow Matching (AAAI 2025 · similarity 0.89)
- SongGen: A Single Stage Auto-regressive Transformer for Text-to-Song Generation (ICML 2025 · similarity 0.88)
- VocalCoachBench: Benchmarking Audio-Language Models on Expert Feedback for Singing (NeurIPS 2026 · similarity 0.88)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.88)
- StyleSinger: Style Transfer for Out-of-Domain Singing Voice Synthesis (AAAI 2024 · similarity 0.86)
- LeVo: High-Quality Song Generation with Multi-Preference Alignment (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.