BAMM
收藏资源简介:
BAMM(Broadcast AI-Music Monitoring)是一个专为真实广播环境中AI生成音乐检测而设计的40小时数据集,由庞培法布拉大学音乐技术组与BMAT公司联合创建。数据集包含约20小时AI生成音乐和20小时人类创作音乐,均来自全球电视广播的原始mp4片段,时长5-60秒,音频采用8kHz单声道低码率编码,保留了广播特有的声学退化与制作上下文。创建过程通过音频指纹技术将清洁参考曲目与广播档案中的出现时间精确匹配,并经过多阶段过滤确保标签可信。该数据集旨在填补现有检测模型在真实广播条件下性能严重下降的研究空白,为评估CNN等检测架构提供了具备实际挑战的基准,推动AI音乐监控技术向工业级应用迈进。
BAMM (Broadcast AI-Music Monitoring) is a 40-hour dataset specifically designed for AI-generated music detection in real-world broadcast environments, co-developed by the Music Technology Group of Pompeu Fabra University and BMAT. The dataset comprises approximately 20 hours of AI-generated music and 20 hours of human-composed music, all sourced from raw MP4 clips of global television broadcasts, with each clip lasting between 5 and 60 seconds. The audio is encoded at 8 kHz in mono with low bitrate, retaining the acoustic degradations and production context unique to broadcast content. During dataset construction, audio fingerprinting technology was employed to precisely align clean reference tracks with their corresponding occurrence timestamps in broadcast archives, and multi-stage filtering was implemented to ensure the credibility of the labels. This dataset aims to fill the critical research gap where current state-of-the-art detection models experience significant performance degradation when deployed in real broadcast scenarios, providing a challenging real-world benchmark for evaluating detection architectures such as CNNs, and promoting the translation of AI music monitoring technology toward industrial-grade practical applications.



