Font Size: a A A

Research On The MIDI Music Retrieval Algorithm Based On Humming

Posted on:2018-02-14Degree:MasterType:Thesis
Country:ChinaCandidate:Y XiaoFull Text:PDF
GTID:2348330518497684Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
As the explosive growth of the music database, the traditional text-based audio retrieval brings greatly inconvenience to the user. The MIDI music retrieval based on the hum is a content-based music retrieval method, the mode allows users to retrieve their need songs just by humming the melody not remembering the lyrics. The goal of this article was to build the complete MIDI audio retrieval algorithm based on humming and verified its feasibility. The main contents of this paper are as follows:1. Audio feature extraction. We analyzed the time domain, frequency domain and cepstrum features of audio signals, introduced some basic expressions of melody contours, and expounded the feature extraction methods of audio signals.2. A humming retrieval algorithm based on HMM. HMM model based on note was constructed to avoid note segmentation. The pitch was converted to the melodic pitch of the melody, thus overcoming differences between singing habits and the different vocal ranges of the humming. The performance of the proposed algorithm was tested with 500 humming test sets, reaching a recognition rate of TOP3 of 78%.3. Humming retrieval algorithm based on depth learning. The key features of each song was obtained by using the 3 layer DBN network structure, which ensured that the melody data could accurately describe the melody of the song, and solved the instability of melody features.Clustering method was adopted to achieve the nearest neighbor retrieval of melody features. 200 Music Libraries of MIDI format were constructed, and the performance of the algorithm was verified by 42 humming query files in wav format, and the recognition rate of TOP3 was 81%. At the same time, a comparison experiment of humming retrieval algorithm based on LSH and humming retrieval algorithm based on DBN was introduced, and the performance of the algorithm based on DBN was proved.The core parts of the two algorithms both included melody feature extraction and melody feature matching, which was the key part of each retrieval algorithm. The melody feature extractions of MIDI music database and technologies related to humming melody feature extraction had been focused on each algorithm.
Keywords/Search Tags:Humming retrieval, Musical Instrument Digital Interface, Melody characteristics, Melody matching, Hidden Markov model, Deep learning
PDF Full Text Request
Related items