## Dataset 

- [UrbanSound8K](https://urbansounddataset.weebly.com/urbansound8k.html)

## Audio files 

Files are converted to melspectrograms that perform better in general for visual transformations of such audio files. 

## Training 

Using With Fast.ai and three epochs with minimal lines of code approaches 95% accuracy with a 20% validation of the entire dataset of 8732 labelled sound excerpts of 10 classes shown above. Fast.ai was used to train this classifier with a Resnet34 vision learner with three epochs.


| epoch	| train_loss |	valid_loss |	accuracy |	time |
|-------|------------|-------------|-------------|-------|
|0	| 1.462791 |	0.710250 |	0.775487 |	01:12 |
| 0	| 0.600056	| 0.309964 |	0.892325 |	00:40 |
|1	| 0.260431	| 0.200901 |	0.945017 |	00:39 |
| 2	| 0.090158 | 	0.164748 |	0.950745 |	00:40 |

## Classical Approaches

[Classical approaches on this dataset as of 2019](https://www.researchgate.net/publication/335862311_Evaluation_of_Classical_Machine_Learning_Techniques_towards_Urban_Sound_Recognition_on_Embedded_Systems)

## State of the Art Approaches 

The state-of-the-art methods for audio classification approach this problem as an image classification task. For such image classification problems from audio samples, [three common](https://scottmduda.medium.com/urban-environmental-audio-classification-using-mel-spectrograms-706ee6f8dcc1)
 transformation approaches are:

- Linear Spectrograms
- Log Spectrograms
- [Mel Spectrograms](https://towardsdatascience.com/audio-deep-learning-made-simple-part-2-why-mel-spectrograms-perform-better-aad889a93505)


## Credits 

Thanks to [Kurian Benoy](https://kurianbenoy.com/) and countless others that generously leave code in github to follow or write blogs that explain various things online. 

## Code Repo & Blog

Additional details on my [Github Repo](https://github.com/gputrain/fastai2-coursework/tree/main/HW) and [my blog](https://www.gputrain.com/) where I will add additional details on this fast ai build, audio transforms and more.