The Study Of Melody Extraction In Query By Humming

Query by humming is a hot spot in the field of research and application of music information retrieval. It usually adopts musical melody information as the feature for the query. The accuracy of the melody feature database directly affects the query results of the query by humming system. Therefore how to extract the standard melody feature is a problem deserving of study.Given the two forms of music including Symbolic data form (MIDI) and waveform data format, this paper studies the melody extraction from two aspects as follows.1. A two-category classification algorithm of MIDI melody channel extraction based on Support Vector Machine is proposed in this paper.In order to ensure the effective extraction of the melody channel, this paper will focus on the musical feature extraction before the classification, the imbalance situation of the samples for classification and the reliability of results after two-category classification in the algorithm. The experimental results show that the proposed algorithm is simple and feasible, meanwhile its accuracy is much higher than the accuracies of other algorithm proposed by related researchers.2. An improved singing voice separation algorithm based on Computational Auditory Scene Analysis is proposed in this paper, which extracts the pitch of singing voice while realizing the melody extraction of waveform music.In the algorithm, the input music signal is preprocessed, and then the singing voice is separated and the pitch of singing voice is extracted based on Computational Auditory Scene Analysis. The main innovative point of this algorithm is the preprocessing of the input music signal with Harmonic/Percussive Sound Separation (HPSS) method before the singing voice separation. This preprocessing could greatly weaken the energy of harmonic component which can seriously affect the separation of singing voice, and make the singing voice separation from the percussive component simple and effective. The experimental results show that the proposed algorithm does not need any prior knowledge, and it could separate the singing voice and extract the pitch effectively.