隨著音樂專輯製作爆炸性的成長，如何管理鉅量的音樂資料以及快速檢索音樂資訊成為一項重要議題。對於鉅量的音樂資料庫，直接從音樂壓縮檔中直接擷取重要的頻率參數來表示音樂特徵，此方式大大的有益於提升音樂檢索速度。此論文中，我們針對高級音訊編碼(AAC)音檔進行分析離散傅立葉轉換(FFT)與離散餘弦轉換(MDCT)之間的頻率轉換差異，並考量轉換後的頻率解析度來選取適當的頻率範圍，進而提出一套在AAC壓縮域中Chroma特徵轉換方法。直接使用AAC壓縮資訊進行Chroma轉換時，其長短窗框轉換機制會致使不同窗框有著不同的頻率解析度，忽略此窗框切換進行Chroma特徵轉換會嚴重的影響其特徵對映的準確性，因此，如何在對有長短窗框切換機制的AAC檔進行Chroma特徵轉換是為一項挑戰。 對於有著較差頻率解析的短窗框，我們提出Peak competition方法合併8個接續的短窗框來增強音調的資訊。而在訊框切割方面，我們提出一簡單動態切割的方法取代複雜度高的節拍追蹤(Beat tracking)。再者，為了能夠處理不同取樣率的AAC音檔，我們提出動態頻率選擇機制來自動選擇不同取樣率以及不同窗框下的頻率範圍。實驗結果顯示，在Covers80資料庫中，我們提出的方法在Top-1音樂搜尋結果比先前壓縮域研究的文獻提升10%準確率，其音樂搜尋效能與現今在原始域的搜尋技術相去不遠，此外，我們所提出的動態頻率選擇方法對於不同取樣率下的AAC檔，其音樂檢索能力呈現穩定且具有相當的準確性。;With the explosive growth in the number of music albums produced, retrieving music information has become a critical aspect of managing music data. Extracting frequency parameters directly from the compressed files to represent music greatly benefits processing speed when working on a large database. In this study, we focused on advanced audio coding (AAC) files and analyzed the disparity in frequency expression between discrete Fourier transform and discrete cosine transform, considered the frequency resolution to select the appropriate frequency range, and developed a direct chroma feature-transformation method in the AAC transform domain. An added challenge to using AAC files directly is long/short window switching, ignoring which may result in inaccurate frequency mapping and inefficient information retrieval. For a short window in particular, we propose a peak-competition method to enhance the pitch information that does not include ambiguous frequency components when combining eight subframes. Moreover, for chroma feature segmentation, we propose a simple dynamic-segmentation method to replace the complex computation of beat tracking. In addition, a dynamic frequency selection method is proposed to deal with various sampling rate of AAC files. Our experimental results show that the proposed method increased the accuracy rate by approximately 10% in Top-1 search results over transform-domain methods described previously and performed nearly as effectively as state-of-the-art waveform-domain approaches did in Covers80 dataset. Furthermore, the proposed dynamic frequency method shows a stable performance for a comprehensive AAC retrieval system.
