本論文主要針對特徵值擷取方法梅爾倒頻譜係數MFCC 中的梅爾濾波器組做研究。 在基於粒子群演算法最佳化濾波器組的中心頻率與邊界頻率上，提出不同於一般使用辨識率當適應函數的方法，而是以統計曲線與濾波器組包絡線的相似度做為適應函數進行最佳化，而本論文依照語音訊號在能量頻譜上的特性，以能量統計圖及能量差異性統計圖為依據，得到兩組最佳化的結果，並分別進行關鍵詞辨識和三種常見雜訊環境下的測試。 最後的實驗結果顯示，此方法有提升特徵值擷取效果的能力，提高了關鍵詞萃取系統的辨識率，且在強健性上亦含有特定環境的抗雜訊能力。 In this thesis, a study for feature extraction using filter bank applied to mel frequency cepstrum coefficients (MFCC) is presented. We propose a novel approach to use particle swarm optimization (PSO) to optimize the parameters of MFCC filterbank, such as the central and side frequencies. The proposed PSO algorithm utilizes filter similarity between statistical curve and filterbank’s envelope as fitness function. According to the energy and energy difference statistical charts that comply with characteristics of the speech signal in the energy spectrum, we obtained two optimal results by PSO. Then keyword recognization and three noisy environments are considered for tests. The results of our experiments show that the proposed method improves the recognition rate of keyword spotting system and the robustness against the testing noisy environments.
