An Efficient Piano Performance Evaluation Model Using DTW based on Deep Learning

A piano performance scoring algorithm based on Convolutional Neural Network (CNN), Dynamic Time Warping (DTW) and Linear Regression are developed. A note-based method is used to make the CNN get a high F1-score so that it can better extract audio features. Combining the actual expert scoring benchmarks with the results from CNN and DTW, a set of more professional indicators, the used linear regression to predict the score and test model on a real database which is provided by a piano association are developed. Finally, the prediction score models based on NN transcription models with different precisions are compared.