調變頻譜正規化法使用於強健語音辨識之研究

Study of Modulation Spectrum Normalization Techniques for Robust Speech Recognition

自動語音辨識在實際系統應用中，語音信號經常受到環境雜訊的影響而降低其辨識率。為了提升系統的效能，許多研究語音辨識的學者歷年來不斷地研究語音的強健技術，期望能達到語音辨識系統的最佳化表現。在本論文中，我們主要是受時間序列結構正規化法觀念所啟發，進而探討並發展出更精確有效的調變頻譜正規化技術。我們提出了三種新方法，包含了等連波時間序列濾波器法、最小平方頻譜擬合法與強度頻譜內插法。這些方法將語音特徵時間序列的功率頻譜密度正規化至一參考的功率頻譜密度，以得到新的語音特徵參數，藉此降低雜訊對語音之影響，進而提升雜訊環境下的語音辨識精確度。同時，我們也將這些新方法結合其他特徵強健化的技術，發現這樣的結合能帶來更顯著之辨識率的提升。

The performance of an automatic speech recognition system is often degraded due to the embedded noise in the processed speech signal. A variety of techniques have been proposed to deal with this problem, and one category of these techniques aims to normalize the temporal statistics of the speech features, which is the main direction of our proposed new approaches here. In this thesis, we propose a series of noise robustness approaches, all of which attempt to normalize the modulation spectrum of speech features. They include equi-ripple temporal filtering (ERTF), least-squares spectrum fitting (LSSF) and magnitude spectrum interpolation (MSI). With these approaches, the mismatch between the modulation spectra for clean and noise-corrupted speech features is reduced, and thus the resulting new features are expected to be more noise-robust. Recognition experiments implemented on Aurora-2 digit database show that the three new approaches effectively improve the recognition accuracy under a wide range of noise-corrupted environment. Moreover, it is also shown that they can be successfully combined with some other noise robustness approaches, like CMVN and MVA, to achieve a more excellent recognition performance.