【摘要】：當前,基于麥克風陣列信號處理的聲源定位技術廣泛應用于各種領域,如視頻會議、語音增強、智能機器人、智能家居等。然而由于各種干擾,會使得聲源定位性能降低,甚至無法定位,特別是室內(nèi)環(huán)境下,常有混響、噪聲等不利因素。因此,對于聲源定位來說,如何能夠提高惡劣條件下的魯棒能力,提升定位準確性是一個研究重點。近年來,基于機器學習算法利用分類識別來進行聲源定位開始得到關注,這類方法比起傳統(tǒng)聲源定位算法不僅有更強的魯棒性,而且能夠在麥克風無法收到直達聲時依舊有效。本文基于機器學習算法研究如何在混響和噪聲環(huán)境下更好地提升室內(nèi)聲源定位的性能。首先分析了聲波傳播模型和麥克風陣列信號接收模型,介紹了傳統(tǒng)的GCC聲源定位算法和SRP-PHAT聲源定位算法,然后簡要總結了機器學習算法。在此基礎上,本文使用相位變換加權廣義互相關函數(shù)作為特征,提出直接使用線性判別分析分類器去識別,仿真結果表明其定位性能在混響嚴重的情況下優(yōu)于樸素貝葉斯分類器。接著利用線性判別分析對互相關函數(shù)進行特征變換,對投影后的特征使用分類識別的方式定位,在惡劣環(huán)境下其定位性能要大大強于未變換前。然后從單一分類器的研究推廣到多個分類器的組合,使用Adaboost和Bagging方法對多個分類器集成,集成后定位性能比單一分類器更好。最后利用優(yōu)化的Bagging方法進行聲源定位,利用K均值聚類方法選擇性集成個體分類器,進一步提高聲源定位的魯棒能力。
[Abstract]:At present, sound source localization technology based on microphone array signal processing is widely used in various fields, such as video conference, speech enhancement, intelligent robot, smart home and so on. However, due to a variety of interference, the performance of sound source location will be reduced, even unable to locate, especially in the indoor environment, there are often adverse factors such as reverberation, noise and so on. Therefore, how to improve the robust ability and improve the accuracy of sound source location under harsh conditions is the focus of research. In recent years, the machine learning algorithm based on classification identification for sound source localization has been paid more attention, this method is not only more robust than the traditional sound source localization algorithm, but also effective when the microphone can not receive direct sound. This paper studies how to improve the performance of indoor sound source location in reverberation and noise environment based on machine learning algorithm. Firstly, the acoustic propagation model and the microphone array signal receiving model are analyzed. The traditional GCC sound source localization algorithm and the SRP-PHAT sound source localization algorithm are introduced. Then the machine learning algorithm is briefly summarized. On this basis, this paper uses the phase transformation weighted generalized cross-correlation function as a feature, and proposes a linear discriminant analysis classifier to identify directly. The simulation results show that the localization performance is better than the naive Bayesian classifier in the case of severe reverberation. Then the linear discriminant analysis is used to transform the cross-correlation function and the projected feature is located by classification and recognition. The localization performance of the projected feature is much better than that before the transformation in bad environment. Then from the study of a single classifier to the combination of multiple classifiers, Adaboost and Bagging methods are used to integrate multiple classifiers, and the performance of the integrated classifier is better than that of a single classifier. Finally, the optimized Bagging method is used to locate the sound source, and the K-means clustering method is used to selectively integrate individual classifiers to further improve the robust ability of sound source location.
【學位授予單位】：南京郵電大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TN912.3;TP181

【參考文獻】

相關期刊論文 前1條

1 萬新旺;吳鎮(zhèn)揚;;基于雙耳互相關函數(shù)的聲源定位算法[J];東南大學學報(自然科學版);2011年05期

，

本文編號：2455768