This paper introduces SoundNet, a deep convolutional network that learns rich natural sound representations from unlabeled video data. The key idea is to leverage the natural synchronization between vision and sound in unlabeled video to transfer discriminative visual knowledge from established visual recognition models into the sound modality. By using two million unlabeled videos, the network learns an acoustic representation that significantly improves performance on standard benchmarks for acoustic scene and object classification. The network is trained without ground truth labels, and visualizations suggest that high-level semantics automatically emerge in the sound network. The SoundNet architecture is a deep convolutional network that processes raw audio waveforms. It is trained by transferring knowledge from vision into sound, and it has no dependence on vision during inference. The network is trained on a large-scale dataset of unlabeled videos collected from the wild, which allows for the training of deeper networks without significant overfitting. The experiments show that the representation learned by the network achieves state-of-the-art accuracy on three standard acoustic scene classification datasets. The paper also explores the use of unlabeled video as a form of weak labeling for audio event classification. It compares the performance of SoundNet with existing state-of-the-art methods and shows that it outperforms them by around 10% on several datasets. The results suggest that unlabeled video is a powerful signal for sound understanding and can be acquired at large enough scales to support training high-capacity deep networks. The paper also presents ablation studies that show the effectiveness of different network depths and teacher networks for visual supervision. The results indicate that deeper networks and stronger vision models can improve sound understanding. The paper concludes that transferring knowledge from vision to sound using unlabeled video is a powerful paradigm for learning sound representations.This paper introduces SoundNet, a deep convolutional network that learns rich natural sound representations from unlabeled video data. The key idea is to leverage the natural synchronization between vision and sound in unlabeled video to transfer discriminative visual knowledge from established visual recognition models into the sound modality. By using two million unlabeled videos, the network learns an acoustic representation that significantly improves performance on standard benchmarks for acoustic scene and object classification. The network is trained without ground truth labels, and visualizations suggest that high-level semantics automatically emerge in the sound network. The SoundNet architecture is a deep convolutional network that processes raw audio waveforms. It is trained by transferring knowledge from vision into sound, and it has no dependence on vision during inference. The network is trained on a large-scale dataset of unlabeled videos collected from the wild, which allows for the training of deeper networks without significant overfitting. The experiments show that the representation learned by the network achieves state-of-the-art accuracy on three standard acoustic scene classification datasets. The paper also explores the use of unlabeled video as a form of weak labeling for audio event classification. It compares the performance of SoundNet with existing state-of-the-art methods and shows that it outperforms them by around 10% on several datasets. The results suggest that unlabeled video is a powerful signal for sound understanding and can be acquired at large enough scales to support training high-capacity deep networks. The paper also presents ablation studies that show the effectiveness of different network depths and teacher networks for visual supervision. The results indicate that deeper networks and stronger vision models can improve sound understanding. The paper concludes that transferring knowledge from vision to sound using unlabeled video is a powerful paradigm for learning sound representations.