A New Approach for Music Generation with LSTM and K-means Clustering Algorithm
A New Approach for Music Generation with LSTM and K-means Clustering Algorithm
Fatemeh Taheri1 Hamed Gorji2
1) Department of Mathematics and Computer science, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran,
2) Master of art in Iranian Instrument performance, Music faculty, University of art, Tehran, Iran,
Publication :
International Conference on Engineering & Technology - Norway (icetconf.com)
Abstract :
A model of music generation with AI needs to recall past details and ability to understand the musical structure. LSMT (Long-short term memory) is possibly the most advantageous type of neural networks and because of the loop in them allowing information to persist, is capable of keeping track of temporally distant events that indicate general music structure. This paper introduces an algorithmic approach by K-means clustering algorithm and LSTM (Long-short time memory) neural network for music generation and discusses its results. First, K-means algorithm for a dataset clustering as an audio processing to recognize the converge input dataset with the maximum similarity based on key audio features which represent a music piece, and LSTM a type of recurrent neural networks to generate a new music content based on the clustering results and music coding structures in MIDI files, combination of clustering method and LSTM performs well in composing music, specifically in range of classical piano music in this study. In fact, this new approach considered audio features and music structures together.
Keywords :
Music Generation
RNN
LSTM
K-means
Audio Features