GTSinger and a multi-technique singing corpus for singing voice research

What the corpus is, and the pairing idea behind it

GTSinger is introduced as a large Global, multi-Technique, free-to-use, high-quality singing corpus with realistic music scores, designed for all singing tasks, along with its benchmarks. It was accepted at NeurIPS 2024 as a Spotlight, and the paper is on arXiv at 2409.13832.

The design decision that makes it different from other singing datasets is the controlled comparison. The key features describe six singing techniques with phoneme-level annotations: mixed voice, falsetto, breathy, pharyngeal, vibrato and glissando. The way the corpus achieves controlled comparison is stated in the data architecture section: each song is divided into comparison groups, a control group which is natural singing without the specific technique, and a technique group which densely employs it.

So for a given song, sung by a given singer, you have two performances, and the only intended difference is the technique. That is what makes technique modelling, technique recognition and technique control trainable. A model that has only one rendition of each song cannot separate the effect of the technique from the effect of the song, the singer and the phrasing, and every downstream task inherits that confusion.

The corpus is also large and recorded properly. 80.59 hours of singing voices are recorded in professional studios by skilled singers, which the documentation claims makes it the largest recorded singing dataset. It is contributed by 20 singers across nine widely spoken languages, Chinese, English, Japanese, Korean, Russian, Spanish, French, German and Italian, and covers all four vocal ranges. The stated purpose of that breadth is to let zero-shot singing voice synthesis and style transfer models learn diverse timbres and styles.

The audience is a researcher, and the release history shows the project still being worked on, with the news entries recording the full dataset release in 2024.09, the code release in 2024.05, the processed data and seven of nine refined languages in 2025.02, and all nine languages processed in 2026.09.

The annotation set, and why realistic scores matter

The annotations are what turn a collection of audio into a corpus, and the README is specific about what is provided.

Alignments and annotations come as TextGrid files, and they include word boundaries, phoneme boundaries, phoneme-level annotations for the six techniques, and global style labels covering singing method, emotion, pace and range. There are also 16.16 hours of paired speech included, which the documentation presents as ensuring comprehensive annotations and broad task suitability.

The paired speech is a quietly important piece. A technique label says something about singing only relative to speech, because the same vocal production in speech and in a song is not the same thing. Paired speech gives a reference point for what the voice does when the singer is not singing at all.

The realistic music scores claim is the other distinctive feature. The documentation contrasts GTSinger with fine-grained music scores, saying that GTSinger features realistic music scores with regular note duration, and that this assists singing models in learning and adapting to real-world musical composition. Scores in musicxml format are provided.

That is a small detail with a large consequence. Human-transcribed scores carry the detail a singer actually follows, including rubato, and a model trained on a transcription that regularises the rhythm learns the score rather than the performance. A corpus whose scores match the performance gives a model something transferable about real singing rather than about notation.

The benchmark directories in the repository root confirm that the corpus comes with the tasks it is meant to support: Style Transfer, Tech-Recognition, Technique-Controllable SVS, STS Conversion and Data-Process. The topic list matches, adding speech-to-sing, style-control, technique-detection and multilingual to the mix.

The directory layout and what the tree tells you

The dataset is organised hierarchically, and the shape is worth reading before downloading anything. There are nine top-level folders, each corresponding to a distinct language. Within each language folder there are five sub-folders, each representing a specific singing technique. And within each technique folder sit the song entries, divided into the control and technique groups.

Here is how the English portion is laid out:

│   ├── EN-Alto-1
│   │   ├── Breathy
│   │   ├── Glissando
│   │   │   └── my love
│   │   │       ├── Control_Group
│   │   │       ├── Glissando_Group
│   │   │       └── Paired_Speech_Group
│   │   ├── Mixed_Voice_and_Falsetto
│   │   ├── Pharyngeal
│   │   └── Vibrato

The naming tells you several things at once. The folder names carry the language and the range, EN-Alto-1, and the range is part of the identifier rather than a separate field. The technique folders are named after the technique, with mixed voice and falsetto combined into a single directory. And the group names use an underscore convention, with Paired_Speech_Group sitting beside the two singing groups, which confirms that the paired speech is organised at the same level as the performances rather than being a separate corpus.

The voice counts are uneven across languages, and the tree shows why this is a designed distribution rather than an accident. English has Alto-1, Alto-2 and Tenor-1, Italian has Bass-1, Bass-2 and Soprano-1, and Korean has Soprano-1, Soprano-2 and Tenor-1. Chinese has Alto-1 and Tenor-1, Russian has only Alto-1, French, German and Japanese have Soprano-1 and Tenor-1 each, and Spanish has Bass-1 and Soprano-1.

So the corpus is deepest where there was most recording time, and the shallowest for Russian and Chinese. Anyone planning a technique experiment on a specific language should check the tree first rather than assume uniform coverage across the nine.

Where the data lives, and the licence split

The download is free and there are two places to get it. The full dataset, which is the audio along with the TextGrid, json and musicxml files, and the processed data, which is metadata.json, phone_set.json and spker_set.json, are both on Hugging Face. There is also a Google Drive link for the same dataset, and a demo page that carries the audio samples and the results of the benchmarks.

What is worth reading carefully is the licence line, because it applies to the data rather than only to the repository. The documentation states that if you are using GTSinger, it means that you have accepted the terms of the licence, linking to dataset_license.md. The repository's own licence field is a custom one that GitHub cannot classify.

So there are two separate things here. The code, the processing scripts and the benchmark implementations sit under the repository licence, while the corpus carries its own terms in a separate file, and using the data is conditioned on accepting them. For a research group that is a normal arrangement, and the term is free-to-use. For a company that intends to train a commercial model on the audio, the distinction between permissive code and conditioned data is the line to read before you build a pipeline around it.

The processing code lives in a Data-Process directory, and the documentation refers to it as the code for processing the dataset, with the repository root holding the Style Transfer, Tech-Recognition, Technique-Controllable SVS, STS Conversion and Data-Process directories. The primary language is Python, the default branch is main, the last push was on 2026-10-06, and the repository is not archived.

What this corpus does not give you

Three limits follow from what the dataset is and how it was built.

The first is the licence constraint, mentioned above and worth repeating because it is the most consequential thing in the readme for a commercial user. The data is free to use and the terms are acceptance-based rather than a permissive licence file, and the repository metadata does not classify the licence either. Anyone planning a product rather than a paper should read dataset_license.md before anything else.

The second is the size of the download and the shape of the files. 80.59 hours of singing at 48kHz with 24-bit resolution in WAV is large even before you add the TextGrid alignments, the json metadata and the musicxml scores, and the paired speech adds 16.16 hours more. The per-singer json file is described as facilitating data parsing and processing for singing models, which is convenient, but the practical reality is that this is a multi-terabyte-adjacent download for anyone working at corpus scale, and a research group will want to think about storage and streaming before committing.

The third is the uneven per-language coverage, which the tree makes visible. The corpus is deepest in English and Italian and thinnest in Russian and Chinese, and while the news entries record that all nine languages are processed as of 2026.09, processed is not the same as equally deep. A technique-detection experiment across nine languages is a different proposition from one across three.

The alternative worth comparing against is a smaller but simpler singing corpus with clean phoneme-level alignment and no technique labels, which is often the better starting point if your task is voice conversion rather than technique control. GTSinger's value is precisely the technique dimension, and if that dimension is not part of your task, you are paying for annotation you do not use.

Editorial conclusion

GTSinger is a good fit for a speech or music technology researcher who needs training data for technique detection, style transfer, singing voice synthesis or style control, because the paired control and technique groups mean a model can learn the difference between two renditions of the same song rather than between two different songs. It is a poor fit if you need a permissive licence for commercial use, because the dataset has its own terms that you accept by using it, and a poor fit if you only need plain singing, because the corpus is organised around technique labels that require the annotation work to be useful. Before downloading, read dataset_license.md, since the terms apply to the data rather than only to the repository code, and plan for the data volume, since 80.59 hours of 48kHz 24-bit audio with TextGrid, json and musicxml is not a small download.

Frequently asked questions

Official sources

1. AaronZ345/GTSinger on GitHub
2. Issues
3. Project website
4. README