Automatic Chord Recognition Based On Hidden Markov Models

With the rapid development of digital music, it is necessary to retrieve music effectively in the digital era. Most existing studies in music retrieve via low-level features such as Mel frequency cepstral coefficient (MFCC) or other spectral coefficients. However, low-level features are insufficient for many applications since they are related to the signal characteristics rather than the semantic content of music. On the contrary, mid-level features such as chord which is the typical characteristic, contain rich information for music analysis and represent musical attributes. Therefore, the mid-level is very useful for musical analysis. Meanwhile, chord have close relationship to potential emotion in music.The same sequence of chords can be found in similar sorts of music, so it can correspondingly be applied to retrivev the similar sorts of music effectively.The purpose of the paper is to construct a complete chord recognition system based on the hidden markov models (HMM). In this paper, the following problems are discussed:1) Extracting the feature of the musical files. The pitch class profile is adopted by analyzing some of the existing expressions of musical files and considering the real-time nature of music retrieval By doing this, the a piece of music can be recognized and expressed more effectively with machines.2) Researching MIDI files and labling in chord. The main type of chord are maior triad, minor triad and diminished triad. The pitchs in an octave below were analyzed and the accordingly maior triad, minor triad and diminished triad were extracted based on pitch. For each chord, the HMM was constructed, respectively.3) Constructing a complete chord recognition system by initializing and training each HMM established. And, finally, the system was detected by testing a MIDI file that is already labeled.In this paper, some good results have been achieved in labeling MIDI corpus. For a paragraph of MIDI file of ten seconds, the label in chord can be extraced correctly. Drawing on the application of HMM in speech recognition, a model for each chord can be constructed correctly.The methods proposed in this thesis can be used in other musical areas. The study of chord recognization are useful for music retrivev , when the recognition rate of chord are high. To some extent, these results verify the good scalability of this system. In addition, we take the pitch class profile as the feature expression of the musical file, which can bring a certain reference value for the analysis of musical emotion.