DiffSinger
openvpi/DiffSinger · Python · Apache-2.0
Diffusion-based singing voice synthesis toolkit
この解説はまだ日本語がありません。英語の原文を表示しています。
DiffSinger is a community SVS stack for research and covers. It sits between TTS and music — useful when you need a singing voice, not just spoken audio.
スコア
導入の要点
git clone https://github.com/openvpi/DiffSinger.git && pip install -r requirements.txt 深読み
DiffSinger is the open-source singing voice synthesis project from OpenVPI: a diffusion-based acoustic model that sings from a musical score with lyrics, pitch and duration — the family of models behind a long line of Chinese vocal-synth communities.
Compared with classic neural SVS it brings cleaner code, adaptive output sampling rate for better sound quality, higher fidelity from improved acoustic models and accelerated diffusion sampling, and more controllability through variance models and parameters.
Pretrained checkpoints, a getting-started guide and configuration docs let you run inference first and train on your own singers when you are ready.