The question was about how to denoise the input. This question is not directly related to the AST model, but I think as for AudioSet, some noises for ASR are actually the recognition target (e.g., vehicle sound, etc) for audio event classification. Therefore, I guess the impact of noise is comparably small for the audio event classification task.
您好，
这个问题和AST模型不是直接相关，我对这个问题了解有限。我认为对于声学事件检测而言，某些在语音识别中被认为的噪声其实是检测目标，并且AudioSet数据本身是来源于用户上传的音频就带有各种环境噪音，因此噪声对声学事件检测任务的影响会较小。如果您指的是通道噪声，可以参阅关于消除通道噪声相关的论文。在语音指令识别的任务中，我们在训练集中人为增加了随机噪声，这一方面可以稍微提高准确率，也应该可能增强模型在噪声环境中的鲁棒性。您可以考虑增加更现实的噪声，比如MS-SNSD噪声数据集。
The question was about how to denoise the input. This question is not directly related to the AST model, but I think as for AudioSet, some noises for ASR are actually the recognition target (e.g., vehicle sound, etc) for audio event classification. Therefore, I guess the impact of noise is comparably small for the audio event classification task.
We did augment random noise for the speechcommands task https://github.com/YuanGongND/ast/blob/87b81895ee866e5be451a5c139751735f836df76/src/dataloader.py#L194, this improves the performance a little bit and might make the model more robust in a noise environment. It might be worth using a more realistic noise dataset such as MS-SNSD.
-Yuan