本文選題：語音增強 + 高階累積量��； 參考：《電子科技大學》2014年碩士論文

【摘要】：語音增強作為通信領(lǐng)域中一項重要技術(shù)手段,隨著通信技術(shù)的快速發(fā)展,它在語音識別、降噪和語音編碼等方面發(fā)揮著越來越重要的作用,成為了近30年中語音處理領(lǐng)域的熱點話題。它以提高語言信號信噪比,改善語音質(zhì)量為目標,進而提高語音信號的舒適度與可懂度,有著十分重要的實用意義。語音信號數(shù)字模型是語音信號處理的基礎(chǔ),模型的準確性將直接影響到語音信號的后續(xù)處理。本文將建立一個全極點模型,該模型結(jié)構(gòu)簡單易于實現(xiàn)。語言增強算法的重點是對語音信號模型和噪聲做參數(shù)估計,模型參數(shù)是卡爾曼濾波算法基礎(chǔ)。噪聲參數(shù)估計主要是通過VAD算法對語音作有效檢測,在信號無聲段通過LPC自相關(guān)法直接估計。接著使用一些常用的參數(shù)估計算法,如極大似然函數(shù)法、Burg算法等,進行仿真實驗。但低信噪比下,這些方法將誤差增大,變的不穩(wěn)定。所以本文在此基礎(chǔ)上應(yīng)用了基于高階累積量的參數(shù)估計算法,根據(jù)其特性可知純凈語音信號與帶噪信號的高階累積量相等。根據(jù)這一事實,我們便可以把AR模型通過高階累積量表示,即MYW方程。而MYW方程的求解方法主要有輔助變量法、LMS算法。通過改進LMS算法,本文提出了共軛梯度算法,通過實驗證明,效果與LMS算法類似,但可以簡化運算。最后通過卡爾曼濾波,實現(xiàn)語音增強。通過仿真實驗結(jié)果分析,高階累積量結(jié)合共軛梯度算法可以很好的實現(xiàn)語言增強,不僅提高了增強算法的適用范圍同時簡化了運算。最后,為了驗證算法的可行性、實時性,本文設(shè)計了硬件平臺;利用芯片TMS320VC5402作處理器,TLV320AIC23B采集音頻信號,STC89LE58RD+作控制器,這樣不僅節(jié)約了成本,同時提高了效率。
[Abstract]:As an important technical means in the field of communication, speech enhancement plays a more and more important role in speech recognition, noise reduction and speech coding with the rapid development of communication technology. It has become a hot topic in the field of speech processing in the past 30 years. It aims at improving the signal-to-noise ratio (SNR) and speech quality of speech signals, and then improves the comfort and intelligibility of speech signals, which is of great practical significance. Speech signal digital model is the basis of speech signal processing, the accuracy of the model will directly affect the subsequent processing of speech signal. In this paper, a full pole model is established, which is simple and easy to implement. The emphasis of speech enhancement algorithm is to estimate the parameters of speech signal model and noise, and the model parameters are the basis of Kalman filter algorithm. Noise parameter estimation mainly uses VAD algorithm to detect speech effectively, and LPC autocorrelation method is used to estimate the noise parameter directly in the silent segment of the signal. Then some commonly used parameter estimation algorithms, such as the maximum likelihood function method and Burg algorithm, are used to carry out simulation experiments. However, at low SNR, these methods will increase the error and become unstable. In this paper, a parameter estimation algorithm based on high order cumulant is applied. According to its characteristics, the high order cumulant of pure speech signal and noisy signal is equal. According to this fact, we can express AR model by higher order cumulant, that is, MYW equation. The main methods for solving MYW equation are the auxiliary variable method and the LMS algorithm. The conjugate gradient algorithm is proposed by improving the LMS algorithm. It is proved by experiments that the effect is similar to that of the LMS algorithm, but the operation can be simplified. Finally, speech enhancement is realized by Kalman filter. The simulation results show that high order cumulant combined with conjugate gradient algorithm can achieve language enhancement, which not only improves the application range of the enhancement algorithm, but also simplifies the operation. Finally, in order to verify the feasibility and real time of the algorithm, this paper designs a hardware platform and uses chip TMS320VC5402 as the controller to collect audio signal from TLV320AIC23B, which not only saves the cost but also improves the efficiency.
【學位授予單位】：電子科技大學
【學位級別】：碩士
【學位授予年份】：2014
【分類號】：TN912.35

【參考文獻】

相關(guān)期刊論文 前5條

1 高鷹,謝勝利;一種變步長LMS自適應(yīng)濾波算法及分析[J];電子學報;2001年08期

2 崔恒志;王，

本文編號：1823022