SINGVERSE
收藏资源简介:
SINGVERSE是一个为唱歌语音增强设计的真实世界基准数据集,它覆盖了多种声学场景,提供了配对的、录音室质量的干净参考录音,用于评估。该数据集包含7942个语音样本,总时长约为18.36小时,跨越19个不同的声学场景,从混响的音乐厅到嘈杂的路边,为稳健的评估提供了丰富的数据。SINGVERSE旨在填补现有数据集在真实世界声学复杂性方面的空白,为唱歌语音增强领域的研究提供了基础性基准。
SINGVERSE is a real-world benchmark dataset designed for singing voice enhancement. It covers diverse acoustic scenarios and provides paired, studio-quality clean reference recordings for evaluation purposes. This dataset consists of 7942 speech samples with a total duration of approximately 18.36 hours, spanning 19 distinct acoustic scenarios ranging from reverberant concert halls to noisy roadside environments, providing abundant data for robust evaluation. SINGVERSE aims to fill the gap in real-world acoustic complexity of existing datasets, serving as a foundational benchmark for research in the field of singing voice enhancement.

- 1SingVERSE: A Diverse, Real-World Benchmark for Singing Voice Enhancement香港中文大学(深圳), 华为中央媒体技术研究院 · 2025年



