KWS, AudioSet Temporally-Strong Labels, Drums
收藏资源简介:
该论文研究了对音频分类模型进行时序定位解释的方法,并提出了一个评估这些解释方法的基准。数据集包括从Librispeech数据集中提取的KWS数据集,用于检测单词"little";AudioSet Temporally-Strong Labels数据集,包含人类标注的语音、音乐和狗吠等事件;以及由鼓声生成的合成数据集,用于检测鼓点。这些数据集均可在论文的网站上访问。
This paper investigates temporal localization explanation methods for audio classification models, and proposes a benchmark for evaluating these explanation methods. The datasets include: 1) the KWS dataset extracted from Librispeech, which is used for detecting the word "little"; 2) the AudioSet Temporally-Strong Labels dataset, which contains human-annotated events such as speech, music, and dog barking; and 3) a synthetic dataset generated from drum sounds, which is designed for detecting drum beats. All these datasets are accessible on the paper's official website.




