On the Improvement of Landmark-based Audio Fingerprinting

• Main Authors: Zenhon Zhuo, 卓真弘
• Other Authors: Jyh-Shing Roger Jang
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/32785416086001379811

碩士 === 國立清華大學 === 資訊工程學系 === 104 === Abstract Audio Fingerprint (AFP) is a fast way of music retrieve. It first records a segment of a music through the microphone on a cellphone or tablet device, and sends the recorded segment to the server for AFP computation. This paper describes several audio fingerprint methods, and put forward improved methods for preprocessing, and using new feature with machine learning to improve the original re-ranking method used. In the preprocessing phase, we set those value in power spectrum with energy smaller than 0 to 0, because these values are not robust to noise, and a high pass filter in the frequency direction added. To improve the original re-ranking method, we extract the new feature with the method proposed by Haistma [8] and compare the similarity of the new feature. Then, using machine learning methods to find a weighted sum of the simlilarities of the new feature and two original features. The machine learning methods used are Pranking, Ranking SVM and Genetic Algorithm. Genetic Algorithm have the best performance among the three methods. The final result is recognition rate raised from 81.21% to 86.04%.