# CNN Architectures for Large-Scale Audio Classification

**Authors:** Shawn Hershey, Sourish Chaudhuri, Daniel P. W. Ellis, Jort F. Gemmeke, Aren Jansen, R. Channing Moore, Manoj Plakal, Devin Platt, Rif A. Saurous, Bryan Seybold, Malcolm Slaney, Ron J. Weiss, Kevin Wilson  
**arXiv:** [1609.09430](https://arxiv.org/abs/1609.09430)  
**Date:** 2016-09-29  
**Summary page:** https://www.summarizepaper.com/en/paper/1609.09430v2/

## Key points

- The paper explores the effectiveness of Convolutional Neural Networks (CNNs) in audio classification.
- Various CNN architectures, including AlexNet, VGG, Inception and ResNet are used to classify a dataset consisting of 70 million training videos with 30,871 video-level labels.
- Analogues of the CNNs used in image classification perform well on their audio classification task.
- Larger training and label sets help up to a point.
- Models using embeddings from these classifiers perform much better than raw features on the Audio Set Acoustic Event Detection (AED) classification task.
- Definitions for mAP (mean average precision), AUC (area under curve), and d-prime metrics are provided.
- Overall, this study demonstrates that CNNs are effective for large-scale audio classification tasks and can be used to improve performance on AED tasks by leveraging embeddings from these classifiers.

## Summary

The paper "CNN Architectures for Large-Scale Audio Classification" explores the effectiveness of Convolutional Neural Networks (CNNs) in audio classification. The authors use various CNN architectures, including AlexNet, VGG, Inception and ResNet to classify the soundtracks of a dataset consisting of 70 million training videos with 30,871 video-level labels. They investigate varying the size of both training set and label vocabulary and find that analogs of the CNNs used in image classification perform well on their audio classification task. The authors also discover that larger training and label sets help up to a point. Furthermore, they compare models using embeddings from these classifiers versus raw features on the Audio Set Acoustic Event Detection (AED) classification task. The model using embeddings from these classifiers performs much better than raw features on this task. The authors provide definitions for mAP (mean average precision), AUC (area under curve), and d-prime metrics and update their results based on changes in the latest Audio Set revision. Overall, this study demonstrates that CNNs are effective for large-scale audio classification tasks and can be used to improve performance on AED tasks by leveraging embeddings from these classifiers.