A Music Cognition Based Music Melody Detection Approach

As the most important expression of music’s motivation and subject, melody is specially studied in field of Music Information Retrieval (MIR), and researchers have made great efforts to find intelligent information processing and analysis methods for melody estimation and analysis. Based on the achievements in music cognition domain from both neuroscience and cognitive psychology, this paper applies the concept of auditory saliency (AS) and proposes a novel approach for melody detection in polyphonic music through the simulation of human’s musical cognition mechanism and characteristics. Firstly in the preprocessing stage, the constant Q transform (CQT) is applied for spectrum calculation, and spectrum model estimation. The AS feature for each semitone is calculated using Bayesian theory according to the semitone’s spectrum distribution over every frequency band. A special time accumulation artificial neural network is used to simulate the human neural system in order to detect salient features as melody candidate contents. In the post processing stage, a novel musicology and cognition related concept of Melody Stream is introduced to regulate melody candidates according to chord perception results. The results of the proposed melody detection methods and its similarity to human perception are evaluated on a small dataset with hundreds of music pieces that cover a number of typical music styles. Experiment results showed that the performance of the proposed strategy may cover more than 75% of the traditional melody line, and the subjective acceptance is measured to more than 90%.