使用MCLT於音訊浮水印之實作與改進

Abstract

音訊浮水印最早用於版權管理，即辨識音訊之所有權。隨著智慧型手機、平板的普及以及運算效能的進步，我們可利用音訊浮水印傳遞資訊。而此項技術最大優點便是硬體需求簡單，只需一擴音器、一麥克風即可。 本論文使用MCLT (Modulated Complex Lapped Transform)來實作音訊浮水印，且利用改變MCLT係數相位藏入資訊。其原理是人類對於相位改變較不敏感，使藏入資訊之音訊與原音訊無異。且MCLT並不會產生blocking artifacts，我們可以得到更好的音訊品質。 音訊浮水印極容易受到干擾，甚至麥克風方向亦會影響辨識率。使用MCLT於音訊浮水印有兩個主要問題，一是有些音訊能量相當小，低能量部份會因此無法抽取資料。二是在各種干擾下，MCLT係數相位發生旋轉，導致抽取資料時發生錯誤。針對第一個問題，本論文提出了將白噪音特定頻帶能量混入原始音訊裡，使能量較低部份能獲得些許能量。第二個問題利用資料分群來解決，本論文嘗試改變K-means 分群的起始中心點獲得改進。 實驗部份本論文以有無加入白噪音、距離、角度、麥克風方向、曲風及片段大小作為實驗變因，嘗試模擬實際應用情形。我們錄製大量音樂訊號送至本論文所提出之系統作辨識，最終可發現加入白噪音能量之音訊在辨識率上獲得極大的進步。

Audio watermark is a technology used for DRM (Digital Rights Management) in earlier days. Now, with the increase of popularity and improvement of computation of smartphones and tablets, we can transmit information via audio watermark. The advantage of audio watermark is that it only requires a speaker and a microphone. In this paper, we implement audio watermark system by using MCLT (Modulated Complex Lapped Transform), and embed data by modifying the phase of the MCLT coefficients because of the imperceptibility of human auditory to modified phase. As a result, we can hardly distinguish the transformed signal from the original audio signal. The MCLT does not produce blocking artifacts so we can get better audio quality. Audio watermark is very sensitive to any acoustic interferences, and even the microphone’s directions will make impact on accuracy. There are two main problems in audio watermark using MCLT. First, some audio signal’s energy is too weak to extract data. Second, the coefficients of MCLT will rotate under some acoustic interferences, and this will lead to data extraction error. For the first problem, we mix the specific frequency band of white noise signal to the audio signal, and increase the energy of weak parts. For the second problem, we use K-means clustering as a solution, and we also try to alter the initial center of K-means clustering to improve the result. In our experiments, the mixing of white noise signals, distances, angles, microphone’s directions, music genre and segment size are independent variables, and we tried many possible combinations to simulate the practical situations. We recorded many audio signals and decoded the result using the proposed system. As a result, we obtain a greater improvement of accuracy by adding white noise signal energy to the audio signal.