非等距麥克風陣列之降噪演算法

Abstract

過去數十年來，利用單麥克風或多麥克風之語音增強演算法持續在不同的訊號處理維度上發展。為了突破單一降噪演算法在效能上的限制，混合式語音增強演算法於近年來陸續被提出。這類混合式語音增強演算法多為結合不同訊號處理維度上的降噪演算法及仰賴精確的雜訊估計。 本論文中，我們提出兩種發展於三顆非等距麥克風陣列之混合式語音增強演算法。這兩種混合式語音增強演算法分別結合兩種發展於能量頻譜密度維度上之降噪演算法，藉此消除方向性及散射白色雜訊構成的混合雜訊。為了消除散射白色雜訊,這兩種混合式演算法皆擁有一個藉由事前雜訊資訊構成的後段濾波器。對於消除方向性干擾源，這兩種混合式系統則分別藉由時間差資訊及兩顆緊鄰麥克風的訊雜比資訊構成另一後段濾波器。我們設計多個客觀評量實驗並通過與多個麥克風陣列之演算法之評比，印證提出的第二種混合式演算法在PESQ及LSD這兩種客觀評量中有最好的表現。從與類似演算法評比中，也印證第一種混合式演算法在結合系統上相對於個別降噪演算法能有更好的表現。從實錄語料的實驗結果也印證提出的兩個系統能夠應用於低迴響的環境。

Over the past few decades, speech enhancement algorithms using a single- microphone or multiple- microphones have been developed in various processing domains. In recent years, hybrid systems for speech enhancement were proposed to overcome the performance limitation of individual noise reduction techniques, which are usually derived in different domains and require the accurate information of the noise field.

In this thesis, we demonstrate two microphone-array algorithms (the proposed system 1 and 2) with non-uniformly distributed three microphones for speech enhancement in directional interferences and diffuse white noise. These two array algorithms are derived in the unified power spectral density domain. Both of them have a post filter based on prior knowledge of the noise field for diffuse noise suppression. To cope with directional interference, the proposed system 1 and 2 have a post filter used interaural time difference (ITD) information and estimates of the signal-to-noise ratio (SNR) at closely spaced dual-microphone respectively. Several experiments are designed to compare the performance of proposed algorithms with performance of other multi-microphone algorithms using objective measures. In terms of objective PESQ and LSD measures, the proposed system 2 with non-uniformly distributed three microphones outperforms the compared algorithms with linear distributed five microphones. Moreover, it has been shown that the cascaded system in the proposed system 1 is more effective in eliminating directional speech interference and diffuse white noise than each individual suppression technique, and it is also the same to compared with a multichannel filtering technique which adjust the de-noising tradeoff between coherent and incoherent noise based on the orthogonal signal decomposition on noise. Furthermore, according to the evaluation results from real recording, the proposed systems have been shown that they are feasible in practical low-reverberant condition.