GTSinger Singing Audio Dataset

This dataset is a global, multi-skill, large-scale open source high-quality singing dataset released by a research team from Zhejiang University in 2024. The relevant paper results are "GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks", has been accepted as a Spotlight in the NeurIPS 2024 Datasets and Benchmarks Track.

The dataset contains 80.59 hours of singing recorded in professional studios by 20 professional singers in 9 different languages, including Chinese, English, Japanese, Korean, etc., providing researchers with a resource library with extremely rich timbres and styles. It is particularly worth mentioning that GTSinger paid special attention to the control and modeling of singing skills during its design, and provided control groups and phoneme-level annotations for 6 commonly used singing skills, which gives it unique advantages in tasks such as singing synthesis and skill recognition.

Another notable feature of GTSinger is that it provides real music scores that match the singing, which is very useful in actual music creation because it is different from fine music scores such as MIDI and is closer to the actual composition process. The structure of the dataset is designed very clearly. Each top-level folder corresponds to a different language, and each language folder is further divided into 5 subfolders, representing specific singing techniques. In addition, the audio quality of GTSinger is very high. All singing and speech are recorded in WAV format at a sampling rate of 48kHz and a resolution of 24 bits, and detailed alignment and annotation information in TextGrid files is provided.

The GTSinger dataset not only excels in data scale and quality, it also supports a variety of singing tasks, including singing synthesis, skill recognition, style transfer, and speech-to-singing conversion, and can be adapted to multiple tasks.