| With the wide application of the Internet, great changes have taken place in the way people access to information. At the same time, it also has a profound impact on people’s way of entertainment. In the network, a lot of music website has provided tens of thousands of pieces of music for the user to choose. In such circumstances, how people from the music library to quickly find their preferred style has become a big challenge. Based on this reality, recommender system based on the user interest in the feeling of the song music through the historical data of users of has been widely studied.This paper studies the realization of a music recommendation system of feature fusion and Gauss mixture model based music. Three important features of music specifically by the wave file of the music: The Gamma-tone cepstrum coefficient(GFCC), Octave-Based Spectral Contrast(OSC), and music mood feature. Using the Gauss mixture model to model the feature vector, By Modulation of the feature vector composed of these characteristics, we can get the long time music features a recommendation list. The main work of this paper is:(1) Use Gamma-tone frequency cepstrum coefficient instead of the Mel-frequency cepstral coefficients to obtain the timbre of music. Gamma-tone cepstrum coefficients are use the Gamma-tone filter group which is more in accordance with human hearing model, and can better simulate the human ear frequency choice model. Through the experimental comparison can be found, compared to the Mel-frequency cepstral coefficient, Gamma-tone frequency cepstral coefficients can produce higher recommendation accuracy.(2) Through the modulation of time axis on the three kinds of feature vectors to get the music long time spectrum characteristics, so it can reflect the overall situation of music. The combination of the long term and short term of features, expresses the basic situation of musical more complete.(3) The use of dynamic K-means algorithm to initialize the Gauss mixture model. Because the number of people liked music style can’t be determined in advance, so we use the K-algorithm to initialize the model of dynamic. Dynamically change K by limiting the size of the cluster radius to get more accurate recommendation.Finally, in order to measure whether the method can produce more accurate recommendation, we established five music database to test. Through the experiment we can draw the conclusion: In the objective aspect, this music recommendation algorithm achieves 87% recommendation accuracy, compared with other similar recommendation algorithm has been improved to some extent; In the subjective aspect, the test also shows that our algorithm is more consistent with human’s true feelings. In a word, the algorithm proposed in this paper effect in the experiment were higher than other algorithm music recommendation system. |
