本論文提出一種新的音訊特徵抽取及音訊辨識的機制。這個機制於小波轉換後的頻率係數上抽取特徵,並且利用秘密分享機制的概念來增強辨識的效能。因此,在不使辨識效能降低的條件下,可以縮短用來做音訊辨識的最短音訊長度。其中我們使用一張二元化的share影像來取代儲存在資料庫中的特徵。此系統是藉由以下三個步驟來辨識一個未知的音訊。1.萃取音訊特徵。2.將此特徵與share影像解碼。3.將解碼出來的結果與一張不變的logo做比對。實驗結果證明此機制是可信賴的並且能抵抗一般的音訊處理。此外,用來做音訊辨識的最短長度可縮短為1.1秒,低於前人之研究。
A novel audio feature extraction and identification scheme is proposed in this thesis. The proposed scheme uses the discrete wavelet transform (DWT) and the concept of secret sharing scheme to improve the robustness and reliability. Hence, the granularity, the minimal length of audio, needed for identification in an audio fingerprinting system, can be reduced without decreasing the efficiency of the system. The scheme employs binary share images to substitute the hash values and the fingerprints stored in the database. The suspect audio signal is then identified by the following steps: 1. Extract the features of the suspect audio. 2. Decode the features with the share images in the database 3. Compare the decoded image to an invariant logo. The experimental results prove the scheme is reliable and robust to some common audio processes. Additionally, the granularity can be reduced to 1.1 seconds, which is less than that of previous work.