audio_tagging

extract features

set the right paths for audio files in the config files first

• python main.py -c config/features.ini
• python main.py -c config/labels.ini

train a modell

after feature extraction

• python main.py -c config/train.ini

Results

Numerous machine learning & signal processing approaches have been evaluated on the ESC-50 dataset. Most of them are listed here. If you know of some other reference, you can message me or open a Pull Request directly.

Terms used in the table:

_{• CNN - Convolutional Neural Network
• LRAP - Label Ranking Average Precision Score}

Title	DataSet	Notes	val_LRAP	Paper	Code
EnvNet (BaseLine)	Mel-spectrogram(train_curated)	CNN + binary_crossentropy probably overfitted, thought the training data is not enough representative	0.5 (77 epoch)	LeCun1998
EnvNet (BaseLine)	Mel-spectrogram(train_curated)+Featurewise center & standardization	CNN + binary_crossentropy probably overfitted, thought the training data is not enough representative	0.51 (31 epoch)	piczak2015b

Requirements

• muda package for data augmentation (pip install muda)

Todos:

benchmark models:

https://github.com/karoldvl/ESC-50

loss for inbalanced classes:

• https://arxiv.org/pdf/1708.02002.pdf http://karol.piczak.com/papers/Piczak2015-ESC-ConvNet.pdf