SVCC2025 挑战赛

当前AI歌声转换（如AI孙燕姿）的核心瓶颈在于"音色可仿，唱法难摹"——形似易而神似难。为攻克这一难题，The Singing Voice Conversion Challenge 2025（SVCC2025挑战赛）正式启动，旨在推动AI语音合成技术发展。

Amphion团队也为SVCC2025挑战赛准备了Baseline: Vevo1.5，一个统一了TTS, VC, SVS, SVC, Editing 等多种生成任务的基座模型，现已开放了所有训练代码、预训练模型。欢迎大家报名参赛！

赛事官网：https://vc-challenge.org/

博客：https://veiled-army-9c5.notion.site/Vevo1-5-1d2ce17b49a280b5b444d3fa2300c93a

赛事任务

Task 1: In-Domain Singing Style Conversion

• Convert source singer A's singing style from style 1 to style 2
• Source singer A is in the training dataset
• Reference singing voice in style 2 from singer A is provided in the training dataset

Task 2: Zero-Shot Singing Style Conversion

• Convert source singer B's singing style from style 1 to style 2
• Source singer B is NOT in the training dataset
• Reference singing voice in style 2 from singer B will not be provided
• Participants would need to use a reference singing voice in style 2 from a different singer in the training dataset to complete the task

赛事数据

Training data

• Contains training data of Task 1 singer A (~4.5 hours, in all 7 singing styles).
• No training data of the Task 2 singer B will be provided.
• Other singers in the training dataset (~70 hours, in all 7 singing styles) will be provided as additional data.
• It will be up to participants how they will choose the target reference style.
• Datasets include waveform files and annotated labels (aligned phoneme and MIDI, global and local style labels, transcriptions).
• The SVCC 2025 dataset is a subset of the GTSinger dataset. Thus, participants will NOT be allowed to use the GTSinger dataset for training. Please refer to the challenge rules for more details.

Test set details

• The participants will be provided with a test set, with each phrase containing 4 source singing styles.
• Participants will then have to convert each phrase into the specified singing styles for each phrase.
• Participants will only be provided with waveform files and NOT the annotated labels.

基线系统

To facilitate the challenge, we will be providing participants with two baseline systems with completely open-sourced codes：

Baseline 1: Serenade

• Paper：https://arxiv.org/abs/2503.12388
• Open-sourced code：https://github.com/lesterphillip/serenade

Baseline 2: Vevo 1.5

• Original paper：https://arxiv.org/abs/2502.07243
• Technical Blog：https://veiled-army-9c5.notion.site/Vevo1-5-1d2ce17b49a280b5b444d3fa2300c93a
• Open-sourced code：https://github.com/open-mmlab/Amphion/tree/main/models/svc/vevosing

重要时间

组织者

• Lester Phillip Violeta, Wen-Chin Huang, and Tomoki Toda (Nagoya University, Japan)
• Xueyao Zhang, Zhizheng Wu (The Chinese University of Hong Kong (Shenzhen), China)
• Jiatong Shi (Carnegie Mellon University, USA)
• Yusuke Yasuda (National Institute of Informatics, Japan)