查詢結果分析
相關文獻
頁籤選單縮合
題 名 | Orthogonal Framework for Linear Prediction Analysis=正交架構下之線性預測分析 |
---|---|
作 者 | 胡懷祖; | 書刊名 | 宜蘭農工學報 |
卷 期 | 16 1998.06[民87.06] |
頁 次 | 頁165-183 |
分類號 | 448.533 |
關鍵詞 | 線性預測分析; 正交轉換; 最小誤差準則; 雜訊敏感度; Linear prediction analysis; Orthogonal transformation; Error minimization criterion; Noise sensitivity; |
語 文 | 英文(English) |
中文摘要 | 在本文中,線性預測是以正交原理加以系統化描述,所形成之彈性架構可方便搭 配不同的最小誤差準則。 針對有聲語音激源的尖突特性, 我們特地採用最小絕對值( LAV )與加權平方數( WLS )來處理時域之線性預測分析。 當被分析之信號呈現於頻域時,我 們亦另行提出兩種方法以達成最大平整頻譜( MSF )要求。 上述四種方法之比較工作是以 合成音共振峰之萃取情況以及對雜訊之敏感度來進行。 所獲致的結論是 LAV 對於所測得的 共振峰相關參數可達成最高的準確度。但是當所採用的激源並非是純粹的脈衝時,LAV 演算 法所需之迭帶次數便會急遽升高。而我們對於雜訊敏感度所做之研究則發現這四種誤差準則 並未明顯優於最小平方準則。 |
英文摘要 | In this paper, the linear prediction is formulated by using the orthogonal principle. The flexibility of the formulated framework facilitates the incorporateion of various error minimization criteria. By exploiting the spiky features of the voiced excitation, we employ the least absolute value (LAV) and the weighted least squares (WLS) to conduct the LP analysis in time domain. While the analyzed signal is presented in frequency domain, two other approaches are proposed to achieve the maximum spectral flatness (MSF). The comparison among the four approaches is demonstrated on the formant extraction from a synthetic vowel with two different excitations and on the sensitivity to additive white Gaussian noise. It is concluded that the LAV achieves the highest accuracy of extracted formant parameters. However, the number of iterations increases dramatically when the adopted excitation is not purely impulsive. Our investigation on the noise sensitivity reveals. that, expect for the MSF in few conditions, all the approaches exhibit no obvious superiority over the least-squares for analyzing noisy speech. |
本系統中英文摘要資訊取自各篇刊載內容。