kayui-gavo/singing-dynamics-restoration

Singing Dynamics Restoration

A synchronized listening demo for wet-only restoration of compressed singing vocals. The public listening set contains all 16 identities in the frozen evaluation population: Chinese, German, Italian, and English, four identities per singer/language stratum. Every example uses the same fixed rb-s001 compression setting; no listening example is selected by restoration quality.

For the current example, Original / Compressed / Proposed are decoded before playback. Switching condition keeps the same timeline and uses only a short click-suppression crossfade. The public Proposed track is the final history-conditioned corrective-gain output. The older intermediate restoration is intentionally omitted from the reader-facing UI.

The exported WAV files are byte-identical copies of the separately materialized presentation assets. Export performs no model inference, checkpoint selection, loudness normalization, gain matching, limiter processing, or scientific metric recomputation.

Audio attribution and license

Listening material is derived from the GTSinger dataset and is published for non-commercial research demonstration under CC BY-NC-SA 4.0.

• —Dataset: https://huggingface.co/datasets/GTSinger/GTSinger
• —Paper: GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks, NeurIPS 2024
• —License: https://creativecommons.org/licenses/by-nc-sa/4.0/