ANN based music chord perception

As a foundation of automatic music transcription, music chord perception is key to analyzing and understanding the structure in computerized music processing and, thus, is the main task of music information retrieval (MIR) systems. This paper describes a chord perception method for polyphonic music based on music cognition theory and artificial neural networks (ANN). The system first applies the constant Q transform (CQT) to the musical piece to get its spectrum. Then, after pitch tuning, onset detection, and pitch estimation which ensure robustness, a pitch class distribution matrix (PCDM) is generated as the feature of the music. An ANN with semi-supervised learning that simulates the human neural system is introduced to complete the chord perception task. Tests show that recognition rate of this strategy is more than 50% which is similar to other systems for polyphonic music with a reasonable computational load.