最小變異數調變頻譜濾波器於強健性語音辨識之研究

A Study of Minimum Variance Modulation Filter for Robust Speech Recognition

本論文所探討的是語音特徵強健性技術，藉此改善雜訊環境下語音辨識的效能。我們利用原始最小變異數調變濾波器法設計的環境失真目標函數，應用至求取濾波器之最佳頻率響應上，進而發展出兩種特徵時間序列濾波器求取演算法，分別為基於最小變異數準則之最小平方頻譜擬合法（MV-LSSF）及基於最小變異數準則之強度頻譜內插法（MV-MSI）。在這兩種方法中，利用我們所求得的濾波器之最佳頻率響應取代原始最小平方頻譜擬合法（LSSF）與強度頻譜內插法（MSI）中所使用的濾波器，來得到欲逼近的目標功率頻譜密度。從Aurora-2連續數字資料庫的實驗結果證實，這兩種基於最小變異數準之調變頻譜正規化法，在各種雜訊環境下都優於傳統的兩種調變頻譜正規化法，而得到更佳的辨識精確度。與基礎實驗結果相比較，MV-LSSF與MV-MSI所達到之相對錯誤降低率分別為在55.41%與51.20%，顯示了我們所提出之新方法能十分有效地提昇語音特徵在雜訊環境下的強健性。

The modulation spectra of speech features are often distorted due to environmental interferences. In order to reduce the distortion, in this paper we apply the minimum variance (MV) criterion to obtain the optimal frequency response of the temporal filter, and then two approaches, least-squares spectral fitting (LSSF) and magnitude spectrum interpolation (MSI) are used to obtain the filtered feature sequence. Accordingly, two new temporal processing approaches are proposed, which are named MV-LSSF and MV-MSI, respectively. In the Aurora-2 clean-condition training task, we show that the new MV-LSSF and MV-MSI give more than 50% relative error rate reduction over the baseline, and provide relative error rate reductions of 8.18% and 2.73% over the conventional LSSF and MSI, respectively. These results reveal that the proposed methods significantly enhance the robustness of speech features in noise-corrupted environments.