This is a tool that uses a convolutional neural network to predict the genre of a given song, provided by the user via a YouTube link. In the demo above, the link is to Chopin's 2nd Nocturne.
It is based on the brilliant work of Valerio Velardo and trained with the MARSAYAS dataset which contains 10 different genres: Jazz, Rock, Disco, Pop, Country, Reggae, Hiphop, Blues, Metal and Classical.
It was built in Python with Tensorflow/Keras and Librosa. The NN architecture is 3 convolutional layers and a dense layer, outputting to a softmax layer as follows:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 128, 11, 32) 320
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 64, 6, 32) 0
_________________________________________________________________
batch_normalization (BatchNo (None, 64, 6, 32) 128
_________________________________________________________________
conv2d_1 (Conv2D) (None, 62, 4, 64) 18496
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 31, 2, 64) 0
_________________________________________________________________
batch_normalization_1 (Batch (None, 31, 2, 64) 256
_________________________________________________________________
conv2d_2 (Conv2D) (None, 30, 1, 64) 16448
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 15, 1, 64) 0
_________________________________________________________________
batch_normalization_2 (Batch (None, 15, 1, 64) 256
_________________________________________________________________
flatten (Flatten) (None, 960) 0
_________________________________________________________________
dense (Dense) (None, 64) 61504
_________________________________________________________________
dropout (Dropout) (None, 64) 0
_________________________________________________________________
dense_1 (Dense) (None, 10) 650
=================================================================
Total params: 98,058
Trainable params: 97,738
Non-trainable params: 320
_________________________________________________________________
If you feed it a generic piece of music from a mainstream western artist, it identifies the genre pretty well. However, the model doesn't generalise exceptionally well, for example classifying most Italian pop as Country. On the test set, the model achieves an accuracy in the order of 77%.
However, it was a tonne of fun to make and I learned a lot, which was the main objective.
The project was built using Test Driven Development, and includes a suite of tests. The API was implemented using Flask.
Install the requirements.
pip install -r requirements.txtRun the server:
python app.pyNavigate to localhost:5000 on your browser.
Enter a valid YouTube link to a musical track <15 minutes long.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.
Shield:
This work is licensed under a Creative Commons Attribution 4.0 International License.
