OpenBMAT (Open Broadcast Media Audio from TV)

Open Broadcast Media Audio from TV (OpenBMAT) 是一个开放的、带注释的数据集，用于音乐检测任务，包含来自 4 个国家的超过 27 小时的电视广播音频，分布在 1647 个一分钟长的摘录中。它旨在包含任何音乐检测数据集的几个基本特征，并且是第一个包含有关音乐响度与其他同时发生的非音乐声音相关的注释。 OpenBMAT 已由 3 个注释者进行交叉注释，获得较高的注释者间一致性百分比，这验证了注释方法并确保了注释的可靠性。

Open Broadcast Media Audio from TV (OpenBMAT) is an open, annotated dataset for music detection tasks. It contains over 27 hours of TV broadcast audio from 4 countries, split into 1647 one-minute excerpts. It aims to include several core characteristics of any music detection dataset, and is the first to contain annotations regarding the correlation between music loudness and other concurrent non-musical sounds. OpenBMAT has been cross-annotated by 3 annotators, achieving a high inter-annotator agreement percentage, which validates the annotation methodology and ensures the reliability of the annotations.

搜集汇总

背景与挑战

背景概述

OpenBMAT是一个用于音乐检测任务的开放数据集，包含来自4个国家超过27小时的电视广播音频，分为1647个一分钟片段。该数据集首次提供了音乐响度相对于非音乐声音的注释，并通过3名注释者的交叉验证确保了高可靠性和一致性。