音乐自动标记测验

当您参加测验时，就会出现答案和解释。

1. 为什么多标签音乐标注器通常使用独立的 sigmoid 输出，而不是对标签使用单一的 softmax？
   • All tag probabilities must sum to one
   • A recording is required to have exactly one tag
   • Sigmoid outputs eliminate the need for labeled examples
   • Several nonexclusive tags can apply to the same recording
2. 大多数现代自动标记模型将什么输入表示输入到其神经网络中？
   • 梅尔频谱图（时频图像）
   • 仅原始波形幅度值
   • MP3 文件的元数据字段
   • 乐谱记谱
3. 为什么使用梅尔标度而不是简单的线性频率标度？
   • 它更有效地压缩文件
   • 它消除了所有背景噪音
   • 它近似人类感知音高的方式
   • 这是版权法所要求的
4. 在现实数据集上训练自动标记模型时的主要挑战是什么？
   • 音频文件太小，无法学习
   • 人工应用的标签有噪音、不完整且不平衡
   • 没有公共音乐数据集
   • 频谱图无法实时计算
5. 哪个数据集通常用于训练和基准音乐自动标记系统？
   • 麦格纳标签音乐
   • 图像网
   • 小队
   • 图书馆演讲
6. What additional search interface can an audio-language model such as CLAP support for a music catalog?
   • Guaranteed identification of every artist without audio input
   • Automatic replacement of all rights metadata
   • Searching only by the recording's compressed file size
   • Retrieval using free-form text descriptions rather than only fixed tag names
7. Which objective commonly trains separate binary tag predictions in a music auto-tagger?
   • A single mutually exclusive genre softmax only
   • Binary cross-entropy over the tag outputs
   • A pixel-reconstruction loss on album artwork only
   • A loss requiring exactly one positive tag per track
8. If a music dataset's annotations are incomplete, how should an absent instrument tag be interpreted?
   • It proves the recording has no instruments
   • It guarantees that the tagger's positive prediction is wrong
   • It may be unannotated rather than reliable evidence that the instrument is absent
   • It means the waveform cannot be used for evaluation