使用不同長度零值嵌入法之語音增強處理技術

Speech Enhancement Using Variable-Length-Zero-Padding Approach

王俊評(Jyun-ping Wang)

指導教授 ： 陸清達

亞洲大學/資訊學院/資訊傳播學系/碩士(2010年)

摘要

語音增強處理是一個被廣泛應用的重要技術，例如：最常見的語音通訊系統和許多不同的語音辨識系統，都可以藉由這項技術的加入而將效能提升；而在語音增強處理的應用中，偵測雜訊頻譜是一項非常具有挑戰性的研究，因為雜訊頻譜估測的好壞，會直接影響增強後的語音訊號品質，所以本論文嘗試在傳送端使用零值嵌入的方式，作為接收端估計通道雜訊的參考位置，使得雜訊估計準確度得以改善，進而提升語音增強的效能。
本論文提出在接收端的信號中，判斷通道雜訊穩定度的方式，作為決定傳送端使用零值音框嵌入法或零值嵌入法的參考依據，達到提昇雜訊估計準確度的目標。當雜訊判斷系統，判斷通道內的雜訊為變化快速的雜訊時，本系統將會自動選用零值音框嵌入法，然後利用零值嵌入的位置，估計通道雜訊頻譜強度，並且搭配頻譜刪減演算法從事雜訊消除處理；相對的，當通道內的雜訊變化速度緩慢時，則系統會自動選用零值嵌入法，然後利用零值嵌入點的位置，估計通道雜訊的取樣點振幅，再搭配取樣點刪減法，從事雜訊消除處理，達到語音增強的目標。
實驗結果證明：本篇論文所提出來的方法，可以較為準確的判斷單通道裡的噪音變化程度，並且選定適合的雜訊估計及語音增強處理方法，改善處理後語音的清晰度，讓接收端可以更清楚的了解發送端所傳送的訊息。

關鍵字

頻譜刪減 ； 零值嵌入 ； 語音增強 ； 雜訊估計 ； 取樣刪減

並列摘要

Speech enhancement is a very important technique in many applications, such as serving as the front-end of the voice communications or the speech recognition systems. It enables the performance of a system to be improved by the application of speech enhancement. The estimation of noise spectrum is still a challenge research for the application of speech enhancement. It is due to the fact that the accuracy of noise estimation can dominate the performance of enhanced speech.
In this thesis , we attempt to analyze the stationary property of a channel noise signal. Hence, the property of noise variation is employed to decide either a zero-padding algorithm or a frame-zero-padding method is performed to improve the accuracy of estimating noise spectral level. In the case of nonstationary noise, the frame-zero-padding approach is employed. The noise spectrum can be estimated during the zero-padded frames. In turn, this noise estimate is applied to a spectral subtraction algorithm for enhancing a noise corrupted speech signal. On the other hand, if the noise is stationary, the zero-padding method is used to estimate the amplitude of noise. This estimate is then applied to a sample thresholding method for speech enhancement.
Experimental results show that the proposed approach can accurately evaluate the stationary property of channel noise. Hence an appropriate zero-padding method is applied in the transmitter. The receiver enhances the corrupted speech in the time domain for stationary noise and in the frequency domain for nonstationary noise. Therefore, the performance of enhanced speech is improved.

並列關鍵字

speech enhancement ； zero-padding ； spectral subtraction

參考文獻

[13]王俊評、陸清達，「使用零值音框嵌入法之語音增強技術」，智慧型系統工程研討會，頁294-297，2010/5。

[17]游明展、王振興，「利用頻譜權重濾波器改善頻譜刪減法於單一通道語音增強」，碩士論文，國立成功大學電機工程學系，台南，2007。

[16]林典蔚、王小川，「語音訊號中的雜訊預估與刪減方法研究」，碩士論文，國立清華大學電機工程學系，新竹，2006。

[4]S. B. Jebara, “A perceptual approach to reduce musical noise phenomenon with Wiener denoising technique,” in Proceedings of International Conference on Aoustics, Speech, and Signal Processing, 2006, pp. 49-52.

[5]L. Singh, S. Sridharan, “Speech enhancement using pre-processing,” in Proceedings of Speech and Image Technologies for Computing and Telecommunications, 1997, pp. 755-758.