以地標為特徵之音訊指紋系統的改進

Abstract

本論文中，我們將針對聲紋辨識 (Audio Fingerprinting, AFP)技術做改進。聲 紋辨識為一種可以快速檢索音樂的技術，主要用途為原曲辨識，使用者可於任何 情況下錄音，以此錄音檔案做為目標，傳入聲紋辨識系統中，最後回報使用者與 此錄音檔最相似之歌曲。聲紋辨識系統亦可用於辨識任何以聲音為傳輸媒介之資 料，例如：演講、電影音訊等。 本論文目標為提升以地標為特徵的音訊聲紋辨識系統，考量地標的能量值， 對抽取地標的步驟做改良，假設能量高的高峰點為越有參考價值的特徵，優先配 對能量高的高峰點，並且給予能量高的高峰點越大的配對數目。再將最終計算地 標分數的權重做調整，結合提出方法得到提高辨識率的改進。

In this thesis, we will improve audio fingerprinting (AFP) systems. Audio fingerprinting is a powerful technology for music retrieval. AFP can be used to identify the most similar audio in the database based on the query. Users can record a query snippet in noisy situation, and input the query snippet to AFP system. AFP system will return the most similar audio. AFP systems can also be used to identify any “sound”, like speech, audio of movie snippets. The goal of this thesis is to improve the accuracy of a AFP system. We improve the step of landmarks extraction by considering the energy of landmarks, and assume that the peaks with higher energy are the more valuable features, thus combining them into landmarks with higher priority. In the final procedure, we calculate the landmark count with a given weight. With all methods, the system achieves 5.29% higher recognition rate than the original AFP system.