加入人工訊號於經驗模態分解法中在語音增強上的研究

Abstract

說話內容常會因背景聲音太大而聽不清楚 ,如何去將語音訊號中所含的噪音清除或抑制,就是所謂的語音增強技術。傳統在單通道語音增強技術中最常採用溫尼濾波器(Wiener filtering)或是頻譜相減法等方法,但大部分均是在頻域上做處理,經過時頻上的轉換,常有語音失真的情形。 黃鍔博士在1998年提出了一種新的訊號分析法希爾伯特-黃轉換(Hilbert Huang Transform, HHT),其方法是將訊號經由經驗模態分解法(Empirical Mode Decomposition, EMD),將資料變化的內部時間尺度作為特徵而分解成多個內建模態函數(Intrinsic Mode Functions, IMF)分量,這些分量經由希爾伯特轉換(Hilbert Huang Transform) 可得到有物理意義的瞬時頻率。近年來經驗模態分解法被應用在語音增強上,針對白噪音分解後的特性,可對各個IMF分量中所含的噪音量做估測並消除。 在本論文我們針對基於經驗模態分解法的語音增強方法作研究。藉由在含噪訊號中加入人工訊號,噪音主要成分在分解過程中將集中在部份分量,移除這些分量以去除大部分噪音,在配合適應性中間值權重濾波器(Adaptive Center Weighted Average filter, ACWA filter)將語音中殘存的噪音消除。實驗顯示,此方法在低訊噪比下有很好的消噪效果,並且可以保留原先的語音特性。

Degradation of the quality of speech caused by the background noise is common in most real situations. How to suppress and remove the noise content in a noisy speech is speech enhancement technique. In traditional signal-channel speech enhancement methods, Wiener filter and spectral subtraction are general methods. But these methods process in frequency domain, the distortion of signal often happen. A new signal analyzing method, Hilbert-Huang Transform (HHT), was proposed by Norden E. Huang et al. in 1998. With EMD, signal can be decomposed into a finite number of intrinsic mode functions (IMFs) based on the local characteristic time scale of the signal. These IMFs with Hilbert transform obtain meaningful instantaneous frequencies. In recent years, EMD was used on speech enhancement. After EMD of white noise, noise component of each IMF can be estimated then remove it. In this thesis, we research on speech enhancement with EMD. After adding an artificial signal to noisy signal, most noise component can concentrate on some IMFs. We can remove most noise by throwing away the IMFs. Adaptive center weighted average filter (ACWA filter) is used to whiten the residual noise in speech. These results of experiment show that the method has good performance of de-noising in low SNR situation and reserve the quality of original speech.