一個針對造假的高解析度音訊的盲檢測演算法

Abstract

隨著高清時代的到來，消費者對於音樂的品質要求也越來越高，其中一點就是音訊解析度的提高。音訊解析度很大程度上是由音訊採樣速率決定的。為了可以辨識高採樣速率音訊是否是由低採樣速率音訊經過重採樣偽造的，本論文提出一個針對造假的高解析度音訊的盲檢測演算法。該演算法分為三個步驟：第一步將已知是否經過重採樣的一組音訊檔經過前處理，包括取差分，取絕對值，取對數和快速傅裡葉轉換等操作；第二步進行取特徵，將資料分為若干子頻段後每個子頻段計算平均值作為特徵值；第三步，將得到的資料分為訓練資料集和測試資料集並輸入SVM分類演算法中訓練出一個模型。這個模型可以判斷後續輸入的音訊檔是否經過重採樣。實驗證明，此演算法對於不同音樂風格、不同採樣速率的音訊檔是否經過重採樣的辨識均有效，且音訊樣本長度僅需100ms便可達到98%以上的準確率。

With the advent of the High-definition age, consumers pay more and more attention to music’s quality, which the most important aspect is the resolution of the audio. Meanwhile, resolution is largely determined by the sampling rate. In order to tell real high sampling rate audio from resampled high sampling rate audio, this thesis proposes a blind detection algorithm for the faked high-resolution audio. The algorithm has 3 steps. First, a pre-processing step, including the process of taking the difference, taking the absolute value, taking the logarithm and taking Fourier transform, to a set of audio which we know if it has been resampled. Secondly, the output data of the first step is divided into several sub-bands, each band take the average as an eigenvalue. Thirdly, the output data of the second step is divided into two groups, training data and testing data, then import them to the SVM algorithm, which can train a model. The model can tell if an audio clip has been resampled. Experiments show that this algorithm is effective to tell if an audio clip has been resampled for different music style as well as different original sampling rate. Moreover, it can reach 98% correct rate when the audio clip is only 100ms.