Research On Music Genre Similarity Detection Algorithm

With the development of Internet and information technology, the number of online music is in the exponential growth. Traditional way to access music can no longer satisfy people’s requirements, which makes us study on music recommendation system that can obtain the list of music automatically.In this master thesis, we present a music similarity detection algorithm based on multi-features and Gaussian mixture model, whose features consist of music timbre, rhythm, and mood. First, we extract gamma tone cepstral coefficients, octave-based spectrum contrast features and emotional features. Then, Gaussian mixture model is used to model these features obtained before. Finally, through similarity measure algorithm, we calculate each model similarity between different songs and obtain the total similarity by a weighted method.The innovation points are as follows:(1) Replace Mel-Frequency Cepstral coefficients with Gamma tone Cepstral coefficients which is an analog of human cochlea characteristics to extract music timbre characteristics. This method has a higher recognition rate and a better robustness.(2) Obtain long-term features by a frame axis modulation technique, which can reduce the dimensions of feature vectors and improve training efficiency. Integrate long-term and short-term features to fully capture the musical characteristics.(3) Initialize Gaussian mixture model parameters with dynamic K-means algorithm to make the model process more accurately.In order to measure this algorithm, five kinds of music database is used to test. Objective test results show that the average recommendation rate is more than86%. The subjective test result shows that the algorithm results agree with human real feelings. Above all, both subjective and objective results are slightly higher than the existing music similarity measure algorithm.