Evaluation of noise-robustness techniques on various types of cepstral features for speech recognition

[[abstract]]在本論文中，我們主要分析介紹應用於語音辨識中之三種語音特徵表示法：梅爾倒頻譜特徵係數(MFCC)、扭曲型傅立葉轉換之倒頻譜係數(WDFTCC)及珈瑪調倒頻譜係數(GTCC)，並凸顯它們之間的差異，另外，我們將三種語音特徵配合數種特徵強健性技術，觀察它們於乾淨環境與雜訊環境下所得到的辨識率，藉此評估比較三種語音特徵其鑑別力、抗噪性及與它們跟強健演算法搭配的加成性，從Aurora-2數字資料庫的實驗評估結果來看，在不使用強健演算法時，傳統的MFCC表現略優於WDFTCC與GTCC，然而當配合強健演算法使用時，WDFTCC與GTCC這兩種新型特徵相較於MFCC能達到更佳的辨識效果。[[abstract]]In this thesis, we first introduce three types of speech feature representation, including mel-frequency cepstral coefficients (MFCC), warped discrete Fourier transform cepstral coefficients (WDFTCC) and gammatone frequency cepstral coefficients (GTCC). These feature representations primarily differ in the applied filter-bank. Then we perform several noise robustness techniques on the above three types of features to evaluate the corresponding additivity in improving the recognition accuracy under a wide range of noise-corrupted environments provided in the Aurora-2 connected digit database. Experimental results reveal that, in the case of no noise robustness processing, MFCC performs slightly better than WDFTCC and GTCC, while GTCC and WDFTCC outperform MFCC when all of them are enhanced by any of the noise robustness approaches in advance.[[note]]碩