Instrument Stream Separation of Single Audio Source Based on Genetic Algorithm

• Main Authors: Kun-Wei Chiu, 邱坤煒
• Other Authors: Tsong-Liang Huang, Ph. D.
• Format: Others
• Language: zh-TW
• Online Access: http://ndltd.ncl.edu.tw/handle/03389799414503919245

碩士 === 國立臺北教育大學 === 資訊科學系碩士班 === 98 === The aim of the thesis is to design a filter system to separate two instruments from a single source melody based on Wiener filter. The system includes two processes : training and separation. In the training process, we extract the feathers of different instruments by their power spectral density, and save them as a database. The feature’s extraction method is based on merge algorithm. In the separation process, the features extracted from training process are applied to linear superposition. By linear superposition, the approximate wave form as the original is extracted. The coefficients of linear superposition are estimated by Genetic Algorithms (GA). Finally, we mix four different instruments’ wave forms and observe its results under different feature’s values and coefficients setting. The results show that if the amount of feature’s values are greater than certain value, the effect on separation quality will go down.