| As a branch of the content-based audio retrieval,humming retrieval has been the concern of many experts and enterprises.Pitch detection,the building of the characteristic and the matching algorithm is the key technology of the humming.About the matching algorithm,this paper makes many improvements,and carry on experiments to verify the effect.Research work are as follows:(1) Based on the idea of system development, the overall design of the system is described, detailed description of the software architecture of the system as a whole, mainly on audio feature extraction module design, humming retrieval module and the database system composition.(2) Mainly to complete the detailed design of the audio feature extraction module. First, investigate the de-noising method for computing the amount of the smaller noise spectral subtraction based on improvements to its gain function, and through experiments show that the improved noise spectral subtraction method can significantly improve the signal to noise ratio; followed by introduction of time-frequency domain analysis, mainly on linear prediction cepstral LPCC and MFCC Mel frequency Cepstral Coefficients and their first-order differential cepstrum parameters, and this article is designed for the platform to explore the use of linear Mel cepstrum(LPC-MFCC) differential cepstrum characteristic parameters as feature extraction methods.(3) Mainly humming retrieval algorithms and software modules designed for use in basic research principles string matching algorithm and HMM algorithm on a continuous hidden Markov model, and pitch shift processing DTW algorithm for improving retrieval accuracy and retrieval speed. Matching algorithms designed for, given the overall framework and the design and development of a prototype system.(4) Software functionality and retrieval algorithms designed in the previous chapters were carried out pilot testing and software testing, simulation experiments, and draw the relevant conclusions. |
PDF Full Text Request