多重正弦激勵的寬頻音訊編碼

Abstract

目前在高品質音訊編碼的研究皆把焦點放在不是次頻帶就是轉換編碼上， 他們皆使用遮蔽效應於動態位元分配中，形成了聽不到的雜訊頻譜。然而 ，這些頻域編碼的方式皆耗費了大量的計算而且多餘的訊息需要進一步消 除。在另一方面，時域編碼若結合一個適當的激發結構，仍有潛力在重現 時達到不失真。在這篇論文中，我們提出一個由不同振幅的正弦函數組成 的多重正弦激發架構，而不是像碼書激發線性預估編碼的白色雜訊架構。 效能結果顯示不僅是在傳統的失真量測如信號雜訊比和分段信號雜訊比， 亦在最新提出的客觀品質量測如一般化的巴克 (bark) 頻譜失真都有顯著 的改善。更進一步地，這多重正弦激發架構亦簡化了90倍的計算量。 Current research on high-quality audio coding focuses on either subband or transform coding whereby the masking effect is exploited in dynamic bit allocation to result in an inaudible noise spectrum. However, the frequency-domain coding approaches consume heavy computations and the redundancy remains to be removed further. On the other hand, time-domain coders have the potential to achieve transparent reproduction by incorporating an appropriate excitation structure. In this thesis, we propose a multi-sinusoids excitation structure which consists of the sum of gain-scaled sinusoidal functions, rather than the white noise as in Code-Excited LPC (CELP). The performance results show significant improvement not only in the conventional distortion measure such as SNR and segSNR, but also in the newly proposed objective quality measure called generalized bark spectral distortion (GBSD). Further more, a multi- sinusoids excitation structure also simplifies the computation load in about 90 times.