Mlamrini/audio_classification_CNN_ESC-50

Audio classification model which uses CNN to train ESC-50 dataset.

Audio Classification with CNN for ESC-50 dataset

Audio classification model which uses CNN to train ESC-50 dataset.

Getting Started

git clone https://github.com/sweat0198/audio_classication_CNN_ESC-50/tree/master

Prerequisites

matplotlib>=2.0.0
pandas>=0.21.1
Keras>=2.1.3
tensorflow>=1.5.0
scipy>=1.0.0
librosa>=0.5.1
scikit_learn>=0.19.1
IPython>=7.8.0
tqdm>=4.48.0
jupyter
notebook

Running

There are simply a jupyter notebook and codes are explained there.

Built With

• Kaggle-notebook - Kaggle notebook used run notebook
• Pycharm - Pycharm is used for working on local machine

Achievements and Outcomes

• Data extended to 5 times with autmentation functions after splitting data as %70 training, %20 validating and %10 testing.
• Model reached to 0.850 accuracy with 0.640 loss in 46 epochs which has 32 batch size.
• Testing made with unseen data by the model, so 0.85 accuracy satisfies the expectations.

Researches

Git Repositories	Comments
karolpiczak/ESC-50	Dataset of 50 classes which each has 40 audio file
shibuiwilliam/audio_classification_keras	Well designed notebook, best accuracy with unseen test files. This is the choosen one.
micah5/pyAudioClassification	Dead Simple Usage, there exist svm, conv1d and conv2d algorithms. To who just want to classify some audio quickly.
seth814/Audio-Classification	With youtube tutorial, it uses Conv1d, Conv2D and LSTM. Could not tested since it cuts audio files.
drscotthawley/panotti	Favorite repo because it has a lot useful tools, but accuracy up to 55 with ESC-50
bheemnitd/Environmental-Sound-Classification	Could not evaluate correctly since it has mixed augmented data in test and train
mtobeiyf/audio-classification	It has also svm, basic neural network and conv1d but accuracies of them are not high enough. Their max accuracy was 0.512
adanRivas/CNN-Audio-Classifier-with-Keras-Tensorflow	65 accuracy with VGG16, pure data, may be it can be retrain with augmented data

Acknowledgments

• https://github.com/shibuiwilliam/audio_classification_keras codes are taken from this repo and some modifications made on it.

Inspirations:

• https://github.com/adanRivas/CNN-Audio-Classifier-with-Keras-Tensorflow
• https://github.com/drscotthawley/panotti

TODO

• Make the prediction flexible in time.