Speech Quality Enhancement by Exploring 1/f Nature of Speech Residual

作     者：Li, Yufeng 

作者单位：Drexel University 

学位级别：M.S., Master of Science/Master of Surgery

导师姓名：Onaral, Banu;Akgiil, Tayfun

授予年度：1998年

主      题：Speech residual Pre-emphasis filters Wavelet coefficients Pitch detection algorithm 

摘      要：Previous work demonstrated the1/f nature of speech residual and proposed a narrowband to wideband speech conversion scheme using this property [6]. This thesis proposes three major improvements of the processing scheme. The residual excited linear predictive (RELP) model is used in this thesis. In this method, the speech is pre-emphasized before linear predictive analysis is performed. The linear prediction coefficients are used to construct the inverse filter for residual extraction. In order to explore the1/f nature of speech for coding and quality enhancement, it is necessary to have good estimation of the spectrum attenuation trend. Thus pre-emphasis filters that can be tuned to different segments of speech are needed. In this thesis, nearly1/f pre-emphasis filters are generated using Park-McClellan method. The proposed processing scheme provides better estimation of the scaling exponent for quality enhancement. The residual of speech is wavelet decomposed for analysis. The pattern of the wavelet coefficients consists of deterministic and random components. In this thesis, deterministic component are separately coded from the random component and reconstruction yield better result. The speech is segmented prior to the analysis. It is well known that since speech is not stationary, an arbitrary segmentation may include both silence and speech, and voiced and unvoiced transition in a given frame. The pitch information can be used to delineate the boundaries between the voiced and unvoiced segments. In this thesis, A pitch detection algorithm that combines the Average Magnitude Difference Function and Sum of Cumulants is proposed. This algorithm outperforms the existing methods in the presence of additive Gaussian noise.