The Music Chords Recognition System Based On Key-Dependent HMM

Chord sequences and chord boundary are the middle-level performance of compactness and robustness of music signal. They have many potential applications, such as music identification, music segmentation, music similarity search and audio compression, and so on. For these reasons, the automatic chord recognition is very attractive for music researchers in the field of information extraction. Subsequently, the researchers also developed a lot of musical chord recognition systems.In recent years, the tremendous success achieved by the Hidden Markov Model (HMM) in speech recognition, has made the researchers begin to think about applying the hidden Markov model to the musical chords recognition. There are already many chord recognition systems to start building a simple hidden Markov model, but the impact of the music tone to the chord recognition was not taken into account. Tone is a very important attribute in music, and sound is generated on the basis of tone. Therefore, the chord tone recognition system should be researched around tone, build Hidden Markov Model based on tone, and add the tone information to the chord recognition system.The main contents of this study are as follows:(1) The introduction part. Firstly, introduce the background of chord recognition system; and then, show some simple achievements of this kind of system; finally, give the organizational structure of our article.(2) The basic theory section. Firstly, introduce explanations of some basic terms in music; secondly, introduce the relevant contents of MIDI documents; thirdly describe the feature vectors of two chords, including chroma vector and tone from the center; finally, detailed show the emergence of HMM, the concept of HMM, the three major problems to be solved by HMM, and the corresponding solution to each problem.(3) The realization part of the system. Firstly, give a overall description of the implementation process of our system; secondly, detailed describe the training process of the system, including the accessing of script files, the accessing of feature vectors, the establishing of hidden Markov model and a key-dependent model, and estimating parameters, including pre-processing of the training data, estimating of the state transition probability, and estimating of observations distribution; finally, give the implementation of viterbi decoding process.(4) The experimental part of the system. According to the implementation process of our system, select training data, establish model based on key-dependent, and estimate the corresponding parameters. Finally, test the system.In this study, the importance of tone is fully considered, introduce the tone information into our musical chords recognition system, and build chord recognition system based on key-dependent HMM.In the process of implementing the system, my main works are as follows:1) Make detailed investigation; acquaint myself with the knowledge development of the chord recognition system; study the content and applications of Hidden Markov Model in depth.2) According to the important influence of tone in the harmonies, introduce the tone information into our musical chords recognition system, and establish HMM model based on the tone identification to identify musical chords. According to music theory, definite 24 tones, and for each tone establish a HMM, and select a maximum possible tone model form the 24 models by Viterbi decoding.3) Construct the flow chart for the realization of the system and the knowledge needed in every realization step, and finally give the detailed design and implementation of our system.4) Give the experiment results of the system. Use MIDI music corpus to train key-dependent HMM, and use 6-dimensional vectors of tone from the center to be as features. Finally a good recognition was gotten by our system.The experiment results showed that the recognition results of Key-dependent model trained in the MIDI music data are better than those of the simple key-independent model. After the above work, our system has basically achieved the desired results, with better recognition. I will continue to learn, and make the corresponding improvement of the software.