A genre-independent chord transcription system from audio using GMM-based HMMs

Hao Wu,Dan Su,Yifang Wang,Xihong Wu
DOI: https://doi.org/10.1007/978-3-642-41407-7_34
2014-01-01
Abstract:Chord transcription is an important task in music processing. In most current implementations, the chord transcription suffers greatly from the high acoustic variance caused by different music styles (such as genre, musician). In this paper, we describe a new style-independent acoustic chord transcription system, which can perform chord transcription on different styles of music directly and gives good frame-level recognition results. In this implementation, all music files are first used without genre clustering to train universal acoustic chord Hidden Markov models, whose state is modeled by single or multiple Gasussian mixture model. Then we extended such model with a probabilistic latent semantic analysis (PLSA) based approach to deal with the acoustic variations. Experimental results show that by the proposed PLSA-based approach, our genre-independent chord transcription system, although recognizing without any genre-specific information of testing data, has outperformed a genre-dependent system. Further analysis is also made to find the most important factors our PLSA-based models have captured. ? Springer-Verlag Berlin Heidelberg 2014.
What problem does this paper attempt to address?