本論文的目的在於改善語音辨識時背景噪音所造成SNR值降低的問題,由於測試環境的不同,語音強化後SNR值高低對於語音辨識模型的匹配會是影響辨識率的主要因素。傳統上,常看到的方法有頻譜刪減法(spectrum subtraction)以及小波濾波(wavelet filter)這兩種方法來濾除雜訊,但濾除效果有限。近年來麥克風陣列方法被提出,語音訊號經由麥克風陣列處理後,可將語音訊號中均值為零雜訊作濾除。因此本論文提出以麥克風陣列為基礎之適應性強化系統,利用前饋式模糊類神經網路訓練頻譜刪減法與小波濾波器彼此間近似最佳化互補比例,並提出分析補償系統(ACS)來補償強化後訊號及雜訊突波處理。在實驗環境中採用Aurora-2語料庫,針對各種加成性噪音做處理,實驗評估後經由我們提出之方法強化過的語音訊號能有效提升SNR值。
In this paper an adaptive filter, based on microphone array technology, is proposed for speech signal enhancement. In the past decade, microphone array technology has been shown to provide effective performance on speech enhancement. In this paper a hybrid filter based on microphone array was proposed to improve the performance on speech enhancement. In the proposed hybrid filter, wavelet filter and spectrum subtraction method were adopted to serve as former processing, then the signals are filtering by microphone array. After that, a Feed-forward Fuzzy Neural Network was employed to obtain the optimal signal mixed ratio between filters. Besides, an analysis compensation system (ACS) was proposed to eliminate the unwanted spur and compensate the filtering signals. Experimental results reveal that our proposed filter outperforms other methods on speech signal enhancement.