本論文提出一個單一麥克風的語音加強法,其目的在於將語音訊號中的外加雜訊去除,以改善語音品質。所提出的方法架構為兩階段的處理方式,首先使用適應性分頻頻譜刪減法將語音中的部分雜訊刪除,然後透過雜訊分類器對預估到的雜訊做傅立葉轉換,觀察其大小在頻域上的分佈,判斷所偵測到的外界雜訊是白色雜訊或有色雜訊,來適當地選擇第二階段的子空間法。若判斷外界雜訊為白色雜訊時,則第二階段將使用具增幅處理的回歸共變異數矩陣子空間法;若判斷外界雜訊為有色雜訊時,將改用一般的回歸共變異數矩陣子空間法來消除雜訊。無論外界雜訊是白色雜訊或有色雜訊,從客觀與主觀測試的結果中得知,我們所提出的語音加強法皆能有效地去除雜訊,來達到改善語音品質的目的。
In the thesis, we proposes a single-microphone speech enhancement algorithm, canceling white noise or colored noise and improving speech quality in noisy speech signal is goal. We propose that framework is two stages. First of all, we use the adaptive frequency separation process of spectral subtraction to filter out segmental noise and as the result of noise classification make fourier transform to observe the noise magnitude distribute situation on frequency domain for estimated noise value, then it will decision the kind of the external noise which is white noise or colored noise and choose the second stage subspace algorithm appropriately. If noise is white, the second stage is recursive covariance matrix subspace which is contained the raise process, else noise is colored, the second stage is a general recursive covariance matrix subspace that they remove noise. However, the external noise is whiten or colored, we understand our proposed algorithm which is effective in canceling noise to improve speech quality on the objective and subjective test simulation.