AaronZ345/GTSinger

What it solves

GTSinger addresses the lack of high-quality, multi-lingual, and multi-technique singing datasets for AI singing models. It provides a comprehensive corpus that allows models to learn not just the melody and lyrics, but also specific singing techniques and realistic musical compositions, bridging the gap between synthetic singing and human-like expression.

How it works

The project provides a large-scale dataset containing over 80 hours of professional studio recordings of 20 singers across nine languages (Chinese, English, Japanese, Korean, Russian, Spanish, French, German, and Italian). The data is organized hierarchically by language and technique, featuring:

• Controlled Comparisons: Audio pairs where a song is sung naturally (control group) and then with a specific technique (technique group).
• Detailed Annotations: Phoneme-level annotations for six techniques (mixed voice, falsetto, breathy, pharyngeal, vibrato, and glissando), global style labels (emotion, pace, range), and manual phoneme-to-audio alignments.
• Realistic Scores: MusicXML files with regular note durations to help models adapt to real-world musical compositions.
• Paired Speech: Over 16 hours of speech data to support speech-to-singing conversion tasks.

Who it’s for

Researchers and developers working on singing voice synthesis (SVS), singing technique recognition, style transfer, and speech-to-singing conversion.

Highlights

• Massive Scale: The largest recorded singing dataset with 80.59 hours of high-quality audio.
• Multi-lingual Support: Covers nine widely spoken languages and all four vocal ranges.
• Technique-Focused: Provides controlled comparisons for six distinct singing techniques.
• Comprehensive Tooling: Includes preprocessing scripts for data segmentation and JSON generation, as well as benchmarks for synthesis, recognition, and style transfer.