Pianalyzer is a simple flask web app that classifies piano music using a midi-note-analyzing convolutional neural network. It is hosted on Heroku, so it might run a bit slow, but it is quite accurate :)
Try out the application here!
- The midi file is converted to JSON by the third party javascript library midi-parser-js
- The JSON is then manually converted into an image with pitch on the y-axis and time on the x-axis
- The image can then be passed through a trained Keras model that classifies the image
- The requests between the front and back end are handled using Flask and Ajax
Chopin's Prelude No. 12
The model uses four 2D convolution layers to recognize broad and specific patterns in the notes such as chords or extensions. To train the model, a moderately small datset of 300 classical pieces and 300 jazz pieces was used. However, the model's validation accuracy was around 99%. In the future, more analysis needs to be done to decide if the model is overfit.